A tunnel ventilation control system and control method based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based tunnel ventilation control system and method, relating to the field of ventilation control technology. The method includes: during the tunnel design phase, establishing an unsteady computational fluid dynamics model based on the tunnel's three-dimensional alignment parameters; calculating the turbulent structure under different traffic conditions using large eddy simulation; and determining the optimal spatial configuration scheme of the fan group based on the velocity and pressure field distributions. During the tunnel construction phase, arranging multiple types of sensor arrays layered along the tunnel arch and sidewalls. During the tunnel operation phase, acquiring real-time vehicle trajectory and speed distribution information from the traffic monitoring system. A dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis is established; a fan collaborative control model considering airflow organization optimization is constructed; and a multi-objective adaptive weight allocation algorithm is used to solve for the optimal operating strategy. When a fire characteristic signal is detected, multiple fan groups are coordinated to form a relay-style smoke exhaust airflow organization.

Description

Technical Field

[0001] This application relates to the field of ventilation control technology, and more specifically to the field of intelligent tunnel ventilation control. In particular, it relates to an artificial intelligence-based tunnel ventilation control system and control method. Background Technology

[0002] During tunnel construction, the ventilation system is a crucial component for ensuring a safe working environment and operational efficiency. Its core function is to promptly remove harmful gases (such as methane and dust) from the construction area, regulate air temperature and humidity, and provide workers with sufficient fresh air, thereby reducing safety risks caused by environmental anomalies and ensuring the smooth progress of construction procedures. Especially in long-distance, high-risk tunnel projects, the effectiveness of ventilation control directly impacts the safety of construction workers and the overall project schedule.

[0003] However, existing tunnel ventilation control methods still have certain limitations. On the one hand, traditional control strategies rely heavily on preset parameters or manual experience for adjustment, making it difficult to respond quickly to complex conditions such as sudden changes in rock mass conditions during construction, which may lead to delayed warnings of harmful gas concentrations. On the other hand, the energy efficiency management of ventilation systems is relatively crude, often failing to dynamically optimize energy allocation according to real-time construction needs, which can easily lead to energy waste and increased construction costs.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides an artificial intelligence-based tunnel ventilation control system and method to solve the above-mentioned technical problems.

[0006] This application provides an artificial intelligence-based tunnel ventilation control system, comprising: a design module for establishing an unsteady computational fluid dynamics model based on the tunnel's three-dimensional alignment parameters during the tunnel design phase, calculating the turbulent structure under different traffic conditions using large eddy simulation, and determining the optimal spatial configuration of the fan group based on the velocity and pressure field distributions; a construction module for layering and arranging multi-type sensor arrays along the tunnel arch and sidewalls during the tunnel construction phase, including a distributed fiber optic temperature sensing system, a miniature pressure sensor network, and a laser particulate matter monitor, to construct a full-section environmental parameter acquisition system; and a data acquisition module for real-time acquisition of temperature gradient field, pressure distribution field, and pollutant concentration field data for each zone through the sensor array during the tunnel operation phase, and simultaneously acquiring real-time vehicle trajectory and speed distribution information from the traffic monitoring system. The system comprises the following modules: a prediction module for establishing a dynamic prediction model of ventilation demand based on spatiotemporal correlation analysis, which analyzes environmental multiphysics data and traffic flow characteristics to predict the migration and diffusion patterns of pollutants in each zone in the future; a solution module for constructing a fan collaborative control model that considers airflow organization optimization, with the goal of minimizing total system energy consumption and the constraint that pollutant concentrations in each zone do not exceed a threshold, using a multi-objective adaptive weight allocation algorithm to solve for the optimal operating strategy; an adjustment module for dynamically adjusting the fan operating status based on the optimization results, prioritizing the activation of downstream fan groups in areas with the largest pollutant concentration gradient to form efficient ventilation corridors; and an optimization module for activating an intelligent graded smoke exhaust mode when fire characteristic signals are detected, dynamically optimizing the smoke exhaust path based on real-time simulation of smoke movement, and coordinating multiple fan groups to form a relay-style smoke exhaust airflow organization.This application provides an artificial intelligence-based tunnel ventilation control method, comprising: during the tunnel design phase, establishing an unsteady computational fluid dynamics model based on the tunnel's three-dimensional alignment parameters, calculating the turbulent structure under different traffic conditions using large eddy simulation, and determining the optimal spatial configuration scheme of the fan group based on the velocity and pressure field distributions; during the tunnel construction phase, layering multiple types of sensor arrays along the tunnel arch and sidewalls, including a distributed fiber optic temperature sensing system, a miniature pressure sensor network, and a laser particulate matter monitor, to construct a full-section environmental parameter acquisition system; during the tunnel operation phase, real-time acquisition of temperature gradient field, pressure distribution field, and pollutant concentration field data for each zone through the sensor array, and simultaneously acquiring real-time vehicle trajectory and speed data from the traffic monitoring system. Information is disseminated; a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis is established, analyzing environmental multiphysics data and traffic flow characteristics to predict the migration and diffusion patterns of pollutants in each zone in the future. A fan collaborative control model considering airflow organization optimization is constructed, with the goal of minimizing total system energy consumption and the constraint that pollutant concentrations in each zone do not exceed thresholds. A multi-objective adaptive weight allocation algorithm is used to solve for the optimal operating strategy. The fan operating status is dynamically adjusted according to the optimization results, prioritizing the activation of downstream fan groups in areas with the largest pollutant concentration gradient to form efficient ventilation corridors. When fire characteristic signals are detected, an intelligent graded smoke exhaust mode is activated, dynamically optimizing the smoke exhaust path based on real-time simulation of smoke movement, and coordinating multiple fan groups to form a relay-style smoke exhaust airflow organization. Based on the embodiments provided in this application, during the tunnel design phase, the turbulent structure and velocity and pressure field distributions under different traffic conditions are accurately calculated using an unsteady computational fluid dynamics model combined with large eddy simulation to determine the optimal spatial configuration of the fan group. Compared to the traditional layout method that relies on experience or simplified models, this approach can avoid problems such as ventilation dead zones and local airflow turbulence caused by unreasonable fan locations from the source, breaking the passive dilemma of "building first and then adjusting" in traditional design and laying the foundation for efficient ventilation in the future. At the same time, during the construction phase, a multi-type sensor array is arranged in layers along the arch and sidewalls to construct a full-section environmental parameter acquisition system. Compared to the traditional situation where the sensor layout is sparse and the parameter acquisition is partial, this approach can comprehensively and in real time obtain temperature gradients, pressure distributions, and pollutant concentration data in various areas of the tunnel, providing complete and accurate basic data support for ventilation control during the operation phase and effectively solving the problem of "delayed early warning of harmful gas concentration due to incomplete parameter acquisition" in the background technology.

[0007] Once operational, the system synchronously collects environmental multi-physics field data and traffic flow information via sensor arrays. Combined with a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis, it can predict the migration and diffusion patterns of pollutants in different zones in advance. This changes the traditional passive control mode of "adjusting only after exceeding standards," preventing pollutants from accumulating and exceeding standards in local areas, and further enhancing the safety of construction personnel. Simultaneously, the constructed fan collaborative control model uses "minimizing total system energy consumption" as the optimization objective and "not exceeding pollutant concentration standards" as the constraint. It solves the optimal strategy through a multi-objective adaptive weight allocation algorithm. Compared to the traditional extensive mode of "emphasizing concentration and neglecting energy consumption" or "blindly starting all fans," it can minimize energy consumption while ensuring pollutant compliance, solving the limitations of "extensive energy efficiency management, energy waste, and high construction costs." Furthermore, it dynamically adjusts the fan operating status based on the optimization results, prioritizing the activation of downstream fan groups in areas with the largest pollutant concentration gradient to form efficient ventilation corridors, avoiding energy waste from "starting all fans in the entire area," and achieving "precise wind control and on-demand energy supply."

[0008] When fire characteristic signals are detected, an intelligent graded smoke extraction mode is activated. Based on real-time simulation of smoke movement, the smoke extraction path is dynamically optimized and multiple fans are coordinated to form a relay-style smoke extraction airflow organization. Compared with the traditional fixed mode of fire smoke extraction (which is difficult to adapt to the dynamic diffusion law of smoke), this method can adjust the smoke extraction strategy according to the real-time movement of smoke, significantly improving smoke extraction efficiency, buying critical time for personnel evacuation and emergency response in the tunnel, further strengthening the safety guarantee capability in tunnel fire scenarios, and comprehensively responding to the core requirement of "ensuring the safety of the construction environment and operational efficiency" in the background technology. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0010] Figure 1 This is a structural diagram of an optional artificial intelligence-based tunnel ventilation control system according to an embodiment of this application;

[0011] Figure 2 A flowchart of an optional artificial intelligence-based tunnel ventilation control method according to an embodiment of this application;

[0012] Figure 3 A flowchart illustrating another optional AI-based tunnel ventilation control method according to an embodiment of this application;

[0013] Figure 4 This is a flowchart of another optional artificial intelligence-based tunnel ventilation control method according to an embodiment of this application.

[0014] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] According to one aspect of the embodiments of this application, an artificial intelligence-based tunnel ventilation control system is also provided. For example... Figure 1 As shown, the system includes:

[0017] Design module 101 is used to establish an unsteady computational fluid dynamics model based on the three-dimensional alignment parameters of the tunnel during the tunnel design phase, calculate the turbulent structure under different traffic conditions using the large eddy simulation method, and determine the optimal spatial configuration scheme of the wind turbine group based on the velocity field and pressure field distribution.

[0018] Construction module 102 is used to arrange multi-type sensor arrays in layers along the tunnel arch and sidewalls during the tunnel construction phase, including a distributed fiber optic temperature sensing system, a miniature air pressure sensor network and a laser particulate matter monitor, to build a full-section environmental parameter acquisition system.

[0019] Data acquisition module 103 is used to collect temperature gradient field, pressure distribution field and pollutant concentration field data of each zone in real time through sensor array during the tunnel operation phase, and simultaneously obtain real-time vehicle trajectory and speed distribution information of traffic monitoring system.

[0020] The prediction module 104 is used to establish a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis. It analyzes environmental multi-physics field data and traffic flow characteristics to predict the migration and diffusion patterns of pollutants in each zone in the future.

[0021] Solver module 105 is used to construct a wind turbine collaborative control model that considers airflow organization optimization. The optimization objective is to minimize the total energy consumption of the system, and the constraint is that the pollutant concentration in each zone does not exceed the threshold. The optimal operation strategy is solved by a multi-objective adaptive weight allocation algorithm.

[0022] The adjustment module 106 is used to dynamically adjust the operating status of the fan according to the optimization results, and prioritize the activation of the downstream fan group in the area with the largest pollutant concentration gradient to form an efficient ventilation corridor.

[0023] The optimization module 107 is used to activate the intelligent graded smoke exhaust mode when a fire characteristic signal is detected. It dynamically optimizes the smoke exhaust path based on real-time simulation of smoke movement and coordinates multiple sets of fans to form a relay-style smoke exhaust airflow organization.

[0024] refer to Figure 1 The tunnel ventilation control based on artificial intelligence in this application relates to intelligent control of tunnel ventilation.

[0025] According to another aspect of the embodiments of this application, such as Figure 2 As shown, this application provides a tunnel ventilation control method based on artificial intelligence, characterized by comprising:

[0026] S201, during the tunnel design phase, an unsteady computational fluid dynamics model is established based on the tunnel's three-dimensional alignment parameters. The turbulent structure under different traffic conditions is calculated using the large eddy simulation method. The optimal spatial configuration scheme of the wind turbine group is determined based on the velocity field and pressure field distribution.

[0027] S202, during the tunnel construction phase, multi-type sensor arrays are arranged in layers along the tunnel arch and sidewalls, including a distributed fiber optic temperature sensing system, a miniature air pressure sensor network and a laser particulate matter monitor, to build a full-section environmental parameter acquisition system.

[0028] S203, during the tunnel operation phase, uses a sensor array to collect real-time data on temperature gradient field, pressure distribution field and pollutant concentration field of each zone, and simultaneously obtains real-time vehicle trajectory and speed distribution information from the traffic monitoring system.

[0029] S204. Establish a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis, analyze environmental multi-physics field data and traffic flow characteristics, and predict the migration and diffusion patterns of pollutants in each zone in the future.

[0030] S205. Construct a wind turbine collaborative control model that considers airflow organization optimization. With the goal of minimizing the total energy consumption of the system and the constraint that the pollutant concentration in each zone does not exceed the threshold, use a multi-objective adaptive weight allocation algorithm to solve the optimal operation strategy.

[0031] S206, dynamically adjust the fan operation status according to the optimization results, and prioritize the activation of downstream fan groups in the area with the largest pollutant concentration gradient to form an efficient ventilation corridor;

[0032] S207: When a fire characteristic signal is detected, the intelligent graded smoke exhaust mode is activated. Based on real-time simulation of smoke movement, the smoke exhaust path is dynamically optimized, and multiple sets of fans are coordinated to form a relay-style smoke exhaust airflow organization.

[0033] It should be explained that this invention provides an intelligent control system for a group of ventilation fans that spans the entire lifecycle of tunnel design, construction, and operation. The core of this method lies in transforming traditional passive-response ventilation into proactive optimization control based on multi-source data fusion and artificial intelligence prediction, ultimately achieving maximum energy savings while ensuring safety.

[0034] During the design phase, determining the optimal spatial configuration is a rigorous process based on accurate physical simulation and multiple iterative optimizations. Specifically, the three-dimensional computational fluid dynamics model not only accurately reproduces the actual geometric contours of the tunnel (including slope, curves, and cross-sectional changes), but also considers in detail the frictional resistance effect caused by the roughness of the inner wall and the complex thermal boundary conditions caused by environmental and vehicle heat dissipation. Using the high-precision numerical method of large eddy simulation, the model can analytically reveal the rapidly changing turbulent vortex structure induced by vehicle motion—a key dynamic characteristic that traditional steady-state simulations cannot capture. Engineers can quantitatively evaluate the merits of different fan layout schemes by analyzing the velocity field isosurface maps (used to identify low-speed recirculation zones or "airflow dead zones") and pressure field contour maps (used to detect abnormally high or low pressure areas) in the simulation results. The optimal spatial configuration is determined through a systematic iterative process: First, an initial fan layout is proposed, simulations are run, and evaluation indicators (such as the uniformity index of wind speed throughout the tunnel, pollutant removal efficiency, and total energy consumption) are calculated. Then, based on flow field analysis, the position, spacing, and jet angle of the fans are adjusted in a targeted manner. For example, the fans are placed in positions that allow their jets to directly cut into the core of the "airflow dead zone," or the angles of adjacent fans are adjusted to allow their jets to superimpose synergistically, so as to effectively break up and suppress large-scale energy dissipation vortices. Finally, among multiple candidate schemes, the scheme with the highest comprehensive performance evaluation score is selected as the final optimal spatial configuration scheme.

[0035] During the construction phase, the deployment of the multi-type sensor arrays follows the principles of hierarchical heterogeneity and redundant reliability. The distributed fiber optic temperature sensing system, acting like a neural network for the tunnel, is laid along the main lines of the tunnel arch and sidewalls, enabling continuous and real-time sensing of the temperature field distribution throughout the tunnel with meter-level spatial resolution. A network of miniature barometric pressure sensors is deployed in high-density arrays at key tunnel sections (such as entrances, exits, slope change points, and areas before and after ventilation fans) to accurately capture pressure distribution and gradient changes along the tunnel, crucial for calculating ventilation dynamics. Laser particulate matter monitors act as precise sentinels, deployed in areas prone to pollutant accumulation (such as the middle of the tunnel and areas with frequent traffic congestion) to accurately monitor the concentrations of PM2.5, PM10, and other particulate matter. All these sensors are integrated and powered through a unified cable network pre-embedded during construction, forming a comprehensive, blind-spot-free, full-section environmental parameter acquisition system, laying a solid data foundation for subsequent intelligent control.

[0036] During the operational phase, in-depth analytics is a crucial data processing step. It goes beyond simply viewing simultaneous, sequential data; instead, it uses algorithms to establish the intrinsic physical correlation between traffic flow and environmental parameters. Specifically, the system's internal processors treat traffic flow data (such as traffic volume, average vehicle speed, and vehicle type composition) as "source terms" driving airflow and pollutant generation within the tunnel. For example, the system identifies dynamic correlation patterns such as "when vehicle speeds on a certain road segment are below 20 km / h and traffic density increases, the CO concentration 500 meters downwind of that segment typically begins to rise significantly after 3 minutes." This coupling relationship is modeled and continuously learned and updated, enabling the system to proactively determine ventilation needs rather than reacting passively.

[0037] The multi-objective adaptive weight allocation algorithm is the core of the system's intelligent decision-making. Its innovation lies in the fact that the weights are not fixed but dynamically adjusted based on real-time risk assessment. The system continuously monitors two key signals: first, the "closeness" of pollutant concentrations in each zone to their safety thresholds; and second, the "acceleration trend" of concentration changes. When the concentration in any zone approaches the limit or shows a rapid upward trend, the algorithm activates a safety-first mode, automatically and significantly reducing the weight of the energy-saving target while increasing the weight of the "ventilation assurance" target, instructing the fan group to strengthen ventilation. Conversely, when the overall air quality is excellent and stable, the system switches to an "energy-efficiency-first" mode, assigning a very high weight to the "energy-saving" target. In this mode, the control system tends to use fewer fans and at lower frequencies to maintain basic ventilation, thereby significantly reducing energy consumption.

[0038] The formation of the high-efficiency ventilation corridor embodies the system's precise spatial air delivery. It refers to the system's use of intelligent algorithms to purposefully select an "optimal airflow path" from the fresh air inlet to the core pollution area within a complex tunnel network. It then coordinates and controls multiple fans along this path, ensuring their jets are precisely synchronized in space and time, forming a concentrated, stable, and powerful "directional airflow band." This "corridor" effectively penetrates dead zones that conventional ventilation struggles to reach, acting like an invisible highway to efficiently and directionally "push" pollutants to the tunnel exit, thus achieving optimal pollutant removal with minimal total energy input.

[0039] The relay-style smoke exhaust airflow organization is an optimized strategy for dealing with special fire conditions. Its core idea is to mimic a "relay race" to achieve efficient transfer and spatiotemporal optimization of smoke exhaust energy. When a fire occurs, the system does not blindly activate all downstream fans simultaneously (which could lead to chaotic airflow and energy cancellation). Instead, it accurately predicts the position and speed of the smoke cloud based on real-time smoke motion simulation. The system first commands the first set of fans closest to the fire source to run at full speed, establishing an initial smoke exhaust airflow with sufficient momentum downstream of the fire source, pushing the smoke cloud to the next section. The system monitors the movement of the smoke front in real time through a sensor network deployed within the tunnel. Once the smoke front has entered and filled the next section, the fans in that section are instructed to start, "taking over" the smoke cloud from the previous set of fans and continuing to push it backward. This process proceeds sequentially, forming an orderly smoke exhaust mode with energy transferred step by step. This method not only avoids the ineffective dissipation of fan energy and ensures the smoothness and reliability of the smoke exhaust process, but also effectively prevents the complex flow field interference that may be caused by all fans starting at the same time near the fire source.

[0040] Furthermore, the establishment of a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis includes:

[0041] Empirical Mode Decomposition (EMD) algorithm was used to process environmental monitoring time-series data to extract intrinsic mode functions of pollutant concentration fluctuations at different time scales.

[0042] A nonlinear mapping network between traffic flow parameters and pollutant diffusion rates is constructed, and multi-step time series prediction is performed using a long short-term memory recurrent neural network.

[0043] Based on the temperature stratification effect and the spatial distribution of longitudinal wind speed, the diffusion coefficient tensor in the pollutant transport equation is dynamically modified.

[0044] The tunnel space is discretized into a structured control volume grid, and the unsteady convection and diffusion equations are solved using the finite volume method to predict the spatiotemporal evolution of pollutants.

[0045] The model parameters and state estimates are updated online by fusing numerical prediction results with real-time monitoring data using an adaptive Kalman filter algorithm.

[0046] It should be explained that the construction of the dynamic prediction model for ventilation demand is a complex, multi-layered, and adaptive process. Its purpose is to improve the traditional prediction based on historical averages to a more accurate and forward-looking prediction based on real-time physical processes.

[0047] Empirical Mode Decomposition (EMD) algorithms play the role of "signal deconstructor" here. The time-series signal of pollutant concentration within a tunnel is typically a complex signal composed of fluctuations generated by various physical causes. The EMD algorithm can adaptively decompose this complex signal into a series of intrinsic mode functions arranged from high to low frequency. For example, it can separate short-duration, high-frequency instantaneous spike fluctuations generated by the instantaneous acceleration of a single large-displacement truck; it can also separate slowly changing low-frequency trend terms that last for several hours, generated by the continuous influx of traffic during the morning rush hour. By analyzing IMFs at different frequencies, the system can more clearly identify the main characteristics and changing patterns of pollutant sources, providing cleaner input features for subsequent predictions.

[0048] Long Short-Term Memory (LSTM) recurrent neural networks are used here as temporal relation learners. Their unique advantage lies in their internal "memory cells," enabling them to learn and remember long-term dependencies within long sequences. During training, the network learns from massive amounts of historical data, thus understanding complex temporal logic such as "Friday evening rush hour congestion patterns differ from weekdays" or "Although traffic flow is currently low, given the continuous increase in traffic over the past two hours, pollutant levels may exceed standards in the next half hour." This makes the model's predictions no longer simple linear extrapolations, but rather possesses "predictive" capabilities based on historical experience.

[0049] Dynamically correcting the diffusion coefficient is a crucial step in ensuring the model closely reflects the real physical world. The diffusion rate of pollutants in a tunnel is not constant; it strongly depends on the local atmospheric conditions. In this invention, the model dynamically selects the diffusion coefficient value that best matches the current operating conditions from a pre-calibrated multidimensional lookup table, calibrated through extensive field experiments and CFD simulations, based on real-time monitored temperature stratification (e.g., the air at the tunnel ceiling is hot and low-density due to heat accumulation, while the road surface is cold and high-density; this stratification inhibits vertical mixing) and longitudinal wind speed (higher wind speeds result in stronger convective transport and altered turbulent diffusion characteristics). This value is then substituted into the pollutant transport control equation for solution. This mechanism significantly improves the model's prediction accuracy and robustness under different seasons, weather conditions, and ventilation conditions.

[0050] The adaptive Kalman filter algorithm serves as a "data fusion and state estimator" here. It not only simply fuses predicted and measured values ​​but also possesses "discrimination" capabilities. The algorithm continuously evaluates the noise statistics of data from each sensor online. If it detects that a CO sensor's reading noise has significantly increased and reliability has decreased due to its late-life stage or temporary contamination, the filter automatically reduces the Kalman gain of that sensor's data during the state update process, thus reducing its influence. Simultaneously, it may correspondingly increase the weight of the model's predicted values ​​or rely more heavily on data from surrounding healthy sensors. This adaptive capability ensures that the system maintains overall prediction stability and reliability even when some sensors experience performance degradation.

[0051] Furthermore, the construction of the nonlinear mapping network between traffic flow parameters and pollutant diffusion rates includes:

[0052] Establish a spatiotemporal graph convolutional network based on an attention mechanism to represent traffic flow data as a spatiotemporal graph structure;

[0053] Vehicle density, average speed, and vehicle type composition features are embedded in the nodes of the graph, and vehicle following relationships and lane change information are embedded in the edges of the graph.

[0054] Local and global spatiotemporal features of traffic flow field are extracted through multi-layer spatiotemporal convolution operations;

[0055] The extracted features are cross-attention calculations performed with pollutant concentration monitoring data in the latent space;

[0056] The time dynamics are captured using a gated loop unit, and the predicted pollutant concentration distribution is output for multiple future time steps.

[0057] It should be explained that the construction of the nonlinear mapping network aims to solve the modeling problem of complex, nonlinear, spatiotemporal correlation between traffic flow and pollutant diffusion.

[0058] The construction of the spatiotemporal graph structure is the foundation for this invention to organically combine physical space with dynamic traffic flow. The specific construction method is as follows: the tunnel is divided along its axial direction into dozens to hundreds of continuous fixed segments, each segment serving as a "graph node." Each node is assigned rich attribute characteristics, including but not limited to: real-time vehicle density, average vehicle speed, and vehicle type composition (e.g., the ratio of cars to trucks). The connections between these nodes are established through "edges." The direction of the edges is usually consistent with the direction of traffic flow, and the edge weights reflect the tightness of the traffic flow connections. For example, weights can be set based on the correlation of traffic flow between upstream and downstream segments, or simply using the reciprocal of the distance between segments as the weight. In this way, the dynamic traffic flow of the entire tunnel is abstracted into a spatiotemporal graph that evolves over time.

[0059] Cross-attention calculation is the "smart eye" that enables models to achieve correlations. When predicting the future pollutant concentration of a target road segment, the model does not consider the influence of all road segments equally. The attention mechanism allows the model to automatically and dynamically calculate an "attention weight" for all other nodes (road segments) in the graph. This weight represents the "importance" of the traffic conditions of other road segments to the pollutant concentration of the target road segment at the current moment. For example, the model may learn that when predicting the concentration of a downstream curved road segment, it needs to pay close attention to the truck density and speed of a long uphill road segment 500 meters upstream, because the emissions from heavy vehicles climbing the slope there are the main source of pollution. Road segments that are closer but have smooth traffic may have a very low attention weight. This cross-attention allows the model to capture complex, long-range causal relationships of pollution transport that go beyond simple spatial proximity, thus making more accurate predictions.

[0060] Furthermore, the step of using a multi-objective adaptive weight allocation algorithm to solve for the optimal operating strategy includes:

[0061] The definition includes a multi-objective optimization function that incorporates energy consumption economic indicators, pollutant distribution uniformity indicators, and airflow stability indicators;

[0062] Based on the real-time environmental characteristics of each partition, the weight coefficient distribution of each indicator in the objective optimization function is dynamically adjusted based on fuzzy inference.

[0063] An improved non-dominated sorting genetic algorithm is used to solve the high-dimensional Pareto optimal solution set, and an elite retention strategy is introduced to maintain population diversity.

[0064] Establish a preference decision-making mechanism based on multi-attribute utility theory to select the optimal control scheme that balances energy saving effect and ventilation quality from the non-dominated solution set;

[0065] The feasible range of key control parameters was determined through global sensitivity analysis.

[0066] Furthermore, the dynamic adjustment of the weight coefficient distribution of each index in the objective optimization function based on fuzzy inference includes:

[0067] The pollutant concentration deviation, concentration change trend, and traffic flow density of each zone are used as fuzzy input variables;

[0068] Design triangular and trapezoidal membership functions to fuzzify the input variables;

[0069] Establish a knowledge base that includes multiple fuzzy rules and define weight allocation strategies for different working conditions;

[0070] The Mamdani fuzzy reasoning method is used for rule reasoning to obtain a fuzzy set of output variables;

[0071] The centroid method is used for defuzzification, and the real-time weight coefficients of each objective function are output.

[0072] It should be explained that the multi-objective optimization solution process is the core of the system seeking the best equilibrium point under multiple constraints.

[0073] The improvements in the improved non-dominated sorting genetic algorithm are mainly reflected in the deep integration with practical engineering constraints. Firstly, regarding chromosome encoding, we abandon the traditional approach of treating wind turbine frequency as a continuous variable, because considering the control precision and equipment lifespan of actual wind turbine inverters, only a limited number of effective speeds are typically set. Therefore, we designed a discrete encoding scheme, where each gene position directly corresponds to the operating state of a wind turbine, such as "00" representing off, "01" representing 30Hz low speed, "10" representing 40Hz medium speed, and "11" representing 50Hz high speed. This makes the search space more realistic, improving search efficiency and the engineering feasibility of the solution. Secondly, in terms of genetic operators, we designed targeted crossover and mutation operations to ensure that the offspring chromosomes still maintain effective discrete encoding. For example, the crossover operation only swaps at the boundaries of the wind turbine group, while the mutation operation randomly switches within the allowed speed range.

[0074] The fuzzy inference system is a logic engine that achieves dynamic adaptive weighting. Its 27 rules are summarized based on domain expert knowledge, extensive historical operational data analysis, and numerical simulation results. To more clearly illustrate its working principle, several representative fuzzy rules are listed here:

[0075] Example Rule 1 (Safety First): IF Concentration Deviation is “High” AND Concentration Change Rate is “Positive” THEN Energy Consumption Weight is “Very Low”, Uniformity Weight is “Very High”. When the concentration in a certain zone has significantly exceeded the standard and is still rising rapidly, it indicates an emergency. At this time, energy consumption should be disregarded, and ventilation should be maximized to dilute the concentration.

[0076] Example Rule 2 (Energy Efficiency Priority): IF Concentration Deviation is “Zero” OR Traffic Flow Density is “Extremely Low” THEN Energy Consumption Weight is “Extremely High”, Uniformity Weight is “Low”. When the overall concentration meets the standard and traffic flow is sparse, the system should reduce its operating power as much as possible to save energy.

[0077] Rule Example 3 (Preventative Adjustment): IF Concentration Deviation is "Low" BUT Traffic Flow Density is "High" THEN Energy Consumption Weight is "Medium" and Uniformity Weight is "Medium". Although the current concentration is not high, the traffic flow is large, indicating that pollutants will accumulate rapidly. The system needs to increase ventilation appropriately in advance to avoid passive response.

[0078] By performing a series of standard steps, such as membership function fuzzification, Mamdani inference, and centroid method defuzzification, these fuzzy rules described in natural language are finally output as accurate, time-varying weight coefficients.

[0079] The preference decision-making mechanism of multi-attribute utility theory is used to make a final choice among numerous Pareto optimal solutions. A Pareto solution set implies that these solutions have varying degrees of advantage in terms of "energy saving" and "ventilation effect," making direct comparison impossible. We pre-define a lexicographical preference for the system: "personnel health and safety" takes absolute priority over "economic operation." During decision-making, the system first examines all Pareto solutions, filtering out those that meet the most basic ventilation safety requirements (e.g., all zone concentrations are below 80% of the safety limit). Then, within this safe solution set, the solution with the lowest total energy consumption is selected as the final implementation plan. This decision-making mechanism ensures that the system's behavior always prioritizes safety.

[0080] Furthermore, the formation of the high-efficiency ventilation corridor includes:

[0081] Gaussian mixture model clustering algorithm is used to identify regions of abnormal concentration of pollutants;

[0082] Calculate the concentration gradient vector field distribution of each zone to determine the main migration direction and diffusion rate of pollutants;

[0083] The optimal ventilation path is planned based on the principle of vector synthesis, and the fan groups downstream of the path are activated first.

[0084] The output power distribution of the fan is adjusted in real time according to the spatiotemporal variation characteristics of the concentration gradient to achieve directional transport and removal of pollutants;

[0085] Establish a dynamic evaluation mechanism for ventilation effectiveness, and verify the effectiveness of the control strategy through the concentration field uniformity index and energy consumption efficiency index.

[0086] Furthermore, the optimal ventilation path planning based on the vector synthesis principle includes:

[0087] The tunnel space is discretized into a directed graph structure, where nodes represent ventilation zones and edges represent potential ventilation paths.

[0088] Define a weight function for each edge, taking into account ventilation resistance, energy consumption cost, and pollutant transport efficiency.

[0089] An improved A* search algorithm is used to find the optimal ventilation path from the pollution source to the outlet in the graph;

[0090] A heuristic function is introduced during the search process to estimate the minimum ventilation cost from the current node to the target node;

[0091] The fan start-up and shutdown sequence on the path is optimized by dynamic programming to ensure a smooth transition in the ventilation process.

[0092] It should be explained that the path planning for forming an efficient ventilation corridor is a dynamic spatial optimization process based on environmental state perception.

[0093] Gaussian Mixture Model (BIC) clustering is used to automatically identify "pollution sources" or "pollution clusters." The number of clusters is not determined subjectively, but rather automatically selected using the Bayesian Information Criterion (BIC), a statistical standard. BIC balances model complexity with good data fit, automatically selecting the most suitable number of clusters, K. This ensures that the algorithm can adaptively identify key control areas regardless of the simplicity or complexity of the pollution distribution within the tunnel.

[0094] The improvement of the A* search algorithm is mainly reflected in the richness of its heuristic function. Traditional A* uses geometric distance as the heuristic cost, while this algorithm uses the following heuristic function: h(n) = α * geometric distance + β * average ventilation resistance + γ * historical pollution frequency. Here, α, β, and γ are adjustment coefficients. Geometric distance ensures optimal spatial length of the path. The average ventilation resistance is calculated based on tunnel design drawings, taking into account factors such as cross-sectional changes, slope, and curvature of curves along the path. This avoids selecting a path that is "short but difficult to ventilate." Historical pollution frequency represents the frequency with which pollution levels have exceeded standards in the historical data of the path area. This guides the algorithm to prioritize paths that "need" cleaning more.

[0095] Through this multi-factor heuristic function, the path searched by the algorithm is not only the shortest in space, but also the one with the highest ventilation efficiency.

[0096] The weighting function is a manifestation of a multi-objective trade-off. Its expression is: Edge Weight = W_r * Ventilation Resistance + W_e * Estimated Energy Consumption - W_t * Transport Efficiency. Here, ventilation resistance is an inherent property; estimated energy consumption is calculated based on the fan performance curve, estimating the power required to generate the desired airflow along the path; transport efficiency is quantified by the sum of current pollutant concentrations in all downstream zones along the path, with a higher value indicating greater pollution removal benefits. The coefficients W_r, W_e, and W_t are not fixed; they are inherited from the fuzzy inference system described above and dynamically adjusted according to the global optimization objective.

[0097] Furthermore, activating the intelligent graded smoke extraction mode includes:

[0098] Multi-source information fusion technology is used to identify the location and scale characteristics of fires, including temperature anomaly distribution detection, smoke concentration gradient analysis, and video image pattern recognition;

[0099] Based on the fire dynamics simulation model, the trajectory of smoke movement is predicted, and the spatiotemporal distribution characteristics of smoke concentration in each zone are calculated.

[0100] Based on flue gas prediction results, smoke emission control zones are dynamically divided, and graded response control strategies are formulated.

[0101] A distributed model predictive control method is used to optimize the start-stop timing and speed ratio of the exhaust fan;

[0102] The pressure field coordinated control algorithm prevents flue gas backflow and fresh air short-circuiting.

[0103] Furthermore, the optimization of the smoke extraction scheme using a distributed model predictive control method includes:

[0104] The tunnel smoke exhaust system is decomposed into multiple mutually coupled sub-control systems;

[0105] A local prediction model is established for each control subsystem to describe its dynamic characteristics and interactions;

[0106] Design a distributed objective function to coordinate the control objectives of each subsystem;

[0107] The alternating direction multiplier method is used to solve distributed optimization problems and achieve coordinated control among subsystems.

[0108] It should be explained that the intelligent graded smoke extraction mode is designed to provide a rapid, effective, and orderly emergency response in the event of a fire.

[0109] The multi-source information fusion method employs a hierarchical decision-making fusion architecture to improve reliability. In the first layer, various sensors perform preliminary intelligent judgments locally: temperature sensors determine if the temperature rise rate exceeds a threshold; smoke sensors determine the increase in smoke concentration; and the video analysis unit identifies flame or smoke textures using computer vision algorithms. In the second layer, each sensor type conducts a vote (e.g., if more than half of the temperature sensors trigger an alarm, the temperature dimension concludes "fire alarm"). In the third layer, a central fusion center makes the final decision based on conclusions from multiple dimensions, including temperature, smoke, and video. We use a weighted voting method, where video image recognition is given the highest weight due to its directness, followed by temperature and smoke. The system only confirms a fire when the weighted score exceeds a set threshold, which significantly reduces system malfunctions caused by false alarms from single sensors.

[0110] The grading criteria for the tiered response control strategy are clearly defined to ensure that the response intensity matches the fire situation.

[0111] Level I Response (Warning Level): Triggered by a single type of sensor (non-video) alarm with low intensity. Actions include: starting 1-2 sets of fans at the nearest downstream end of the fire source at low speed to enhance ventilation in the area, increasing monitoring frequency, and notifying management personnel for confirmation.

[0112] Level II Response (Confirmation Level): Triggered by alarms from at least two different types of sensors or video confirmation of visible smoke. Action: Initiate relay-style smoke extraction mode and sequentially start downstream fans as described above.

[0113] Level III Response (Severe): Triggered by a strong alarm from all types of sensors or video confirmation of an open flame. Action: Building upon Level II response, implement "upstream negative pressure closure," which involves adjusting the two nearest fans upstream of the fire source to operate in a special "high negative pressure mode." This creates a stable negative pressure zone near the fire source with an intensity no less than -10 Pa, acting like an airlock to effectively prevent fresh air from flowing into the combustion chamber and smoke from spreading upstream.

[0114] The specific implementation process of distributed model predictive control is a typical "decomposition-coordination" process:

[0115] System decomposition: The long tunnel is decomposed into N subsystems according to physical fire protection zones or fan control range.

[0116] Local prediction and optimization: In each control cycle, each subsystem i performs the following operations in parallel:

[0117] Based on its own local model (such as a simplified one-dimensional convection-diffusion equation), the key states of the subsystem in the future (such as smoke concentration and wind speed) are predicted.

[0118] Calculate a locally optimal control sequence to minimize the local objective function (e.g., reduce the smoke concentration in the area as quickly as possible).

[0119] The key is that, during the calculation, it must take into account the "coupling information" from the adjacent subsystem j. For example, subsystem i will receive a prediction from subsystem i-1 (upstream): "The smoke from here will reach you in 1 minute," and a request from subsystem i+1 (downstream): "I am currently under high exhaust pressure and would like you to increase the airflow to support me."

[0120] Global Coordination: All subsystems submit their calculated preliminary control schemes and coupling boundary predictions to a coordinator. The coordinator checks whether these schemes are consistent and optimal at the global level. If they are inconsistent, the coordinator generates a set of coordination signals (mainly Lagrange multiplier updates for the coupling boundary states) using the alternating direction multiplier method and broadcasts them to all subsystems.

[0121] Iterative convergence: Each subsystem re-solves its local optimization problem based on the new coordination signal. This process is repeated iteratively until all subsystems reach a consensus on the coupling boundary and the global objective cannot be further improved. At this point, each subsystem executes the first step of its optimal control sequence. In the next control cycle, the entire process is repeated to achieve rolling optimization. This method transforms a complex large-system optimization problem into multiple smaller, parallel-solvable subproblems, and ensures global optimality through coordination, making it very suitable for distributed control architectures in long tunnels.

[0122] According to another aspect of the embodiments of this application, such as Figure 3 As shown, this application provides an artificial intelligence-based tunnel ventilation control method, including:

[0123] S301, deploy a strain monitoring network in the tunnel construction area to collect micro-strain data of the rock mass in real time;

[0124] S302: When low-frequency strain fluctuations that conform to the characteristics of rock mass fracture precursors are detected, a stress early warning signal is generated.

[0125] S303 inputs rock mass microstrain data and stress early warning signals into the strain energy and gas release correlation model, and outputs gas risk level and ventilation control instructions;

[0126] S304, Execute ventilation control instructions, including: activating protective equipment according to the gas risk level, and adjusting the main fan operating status according to the instruction parameters;

[0127] In some embodiments, the command parameters of the ventilation control command refer to the specific setting values ​​used to adjust the operating status of the main fan in the ventilation control command. These parameters need to be dynamically determined according to the gas risk level to ensure that the fan operation matches the risk. Command parameters include, but are not limited to, the operating frequency of the main fan (e.g., 50Hz, 30Hz), output wind speed (e.g., 15m / s, 8m / s), wind pressure threshold (e.g., 200Pa, 100Pa), and duct switching time (e.g., immediate switching, switching after 5 minutes). For example, when the gas risk level is "high", the command parameters may be "frequency 50Hz, wind speed 15m / s, wind pressure 200Pa" to ensure rapid gas discharge; when the risk level is "low", the parameters may be "frequency 30Hz, wind speed 8m / s" to balance safety and energy efficiency.

[0128] In practice, a dynamic mapping rule can be established to link the three levels of gas risk with protective equipment: Level 1 risk activates the automatic spray dust suppression system around the tunnel face; Level 2 risk additionally activates the water supply valve of the explosion-proof water canopy in the ventilation roadway; Level 3 risk simultaneously triggers the emergency evacuation broadcast of the personnel positioning system.

[0129] Design a main fan response algorithm based on rock mass fracture characteristics: Analyze the waveform distortion index parameter in the ventilation control command; when the index exceeds the preset distortion threshold, activate the main fan's high-frequency pulsation mode: use twice the frequency of the rock mass micro-strain fluctuation as the base speed, and superimpose a sinusoidal disturbance with amplitude increasing with the distortion index; monitor the return air gas concentration fluctuation phase in real time under pulsation mode; when the gas concentration phase lags the fan speed phase by more than a preset angle, automatically switch to fixed-frequency boost mode;

[0130] Implement a local ventilation compensation mechanism: In a level 2 or higher risk state, activate a local fan coordination protocol triggered by a stress warning signal; use the sensor unit that generates the warning signal as the origin of the coordinate system to calculate the spatial topology weights of all local fans within a 30-meter radius; the weight allocation adopts a stress wave attenuation model: for every 10-meter increase in distance from the origin, the weight value decreases at an exponential decay rate; dynamically increase the target fan's speed reference value according to the weight ratio.

[0131] Perform closed-loop verification of equipment status: After each start-up of the protective equipment, verify the uniformity of water pressure distribution in the spray system through the pressure sensor array; if the water pressure gradient is detected to exceed the preset differential pressure threshold, switch to the backup water circuit and trigger a maintenance alarm; after the main fan operating parameters are adjusted, collect the motor vibration spectrum for resonance point scanning; when vibration energy is found to be concentrated at a specific frequency, automatically insert a band-stop filter to suppress mechanical resonance.

[0132] S305, real-time measurement of energy input and metabolic output of ventilation system, and calculation of ventilation energy efficiency index as the ratio of metabolic output to energy input;

[0133] S306 dynamically adjusts the fan control parameters and regional power distribution ratio based on real-time construction progress information, gas risk level, and ventilation energy efficiency indicators.

[0134] Optionally, the maximum ventilation mode can be implemented when the predicted gas concentration exceeds a preset safety threshold, and the energy efficiency requirements can be temporarily relaxed after a specific construction phase.

[0135] Furthermore, the strain monitoring network is a distributed optical fiber sensing network, which includes corrosion-resistant armored optical fibers laid in a serpentine pattern along the surface of the tunnel's initial support. The optical fiber diameter is 0.25 mm, and a sensing unit is set up every 10 meters.

[0136] It should be noted that the initial support, or tunnel initial support, is a temporary or permanent support structure (usually including shotcrete, steel arches, anchor bolts, etc.) that is immediately constructed after tunnel excavation to prevent the surrounding rock from collapsing. The initial support surface refers to the outer surface of the initial support structure that is in contact with the surrounding rock or the surface facing the inside of the tunnel (such as the surface of the shotcrete layer).

[0137] Serpentine laying refers to laying optical fibers along the surface of the tunnel's initial support in a curved, continuous, zigzag pattern (similar to the crawling trajectory of a snake), rather than in a straight line. Due to unevenness on the tunnel's initial support surface caused by steel arches, anchor bolts, etc., serpentine laying allows the optical fibers to fit more closely to the surface, reducing uneven stress on the fibers caused by protrusions in the support structure, and ensuring more sensitive monitoring of micro-strain in the rock mass. Armoring refers to the outer layer of the optical fiber being wrapped in a metal protective sheath (such as stainless steel) to resist mechanical wear; corrosion resistance means that the optical fiber and armor material can withstand corrosion from moisture, dust, and small amounts of acidic / alkaline gases (such as hydrogen sulfide generated during construction) within the tunnel. When low-frequency strain fluctuations consistent with precursory characteristics of rock mass fracture are detected, a stress warning signal is generated, including:

[0138] The collected rock mass microstrain data was decomposed into three independent frequency bands for processing. The first frequency band is the range of 0.1Hz to 1Hz, the second frequency band is the range of 1Hz to 10Hz, and the third frequency band is the full frequency band.

[0139] Low-frequency strain fluctuations that conform to the characteristics of rock mass fracture precursors are identified in real time in the first frequency band and used as core triggering events;

[0140] Among them, the precursory characteristics of rock mass fracturing refer to the detectable abnormal features caused by changes in internal stress before obvious fracturing (such as fissure propagation or collapse) occurs in the rock mass, which are key signals for judging the stability of the rock mass. For example, the energy of low-frequency (0.1-1Hz) strain fluctuations accumulates continuously (e.g., the original fluctuation amplitude ≤5με suddenly increases to ≥10με and lasts for more than 30 seconds); the strain curve shows a "step-like rise" (a small jump every 10 seconds, reflecting the gradual propagation of fissures); the periodicity of strain fluctuations increases (e.g., a peak occurs every 5 seconds, reflecting the periodic stress on the rock mass).

[0141] When a core triggering event is detected, a multi-band comprehensive evaluation is performed, including: the first processing channel calculates the cumulative energy value of the first frequency band and generates energy characteristic quantities through integration via a time window; the second processing channel identifies the waveform distortion characteristics of the second frequency band and detects the pulse asymmetry index; the third processing channel analyzes the autocorrelation attenuation characteristics of the third frequency band and evaluates the stability index of the rock mass structure.

[0142] The time window refers to a set continuous time period (such as 10 seconds), and the integral refers to the cumulative calculation of the strain data of the first frequency band within the time window. The result is the energy characteristic quantity, which reflects the degree of energy accumulation of rock mass strain during this period.

[0143] Waveform distortion refers to the deviation of the strain signal waveform in the second frequency band from the normal symmetrical shape (e.g., the amplitudes of positive and negative pulses are equal in a normal waveform, but not in a distorted waveform); the pulse asymmetry index refers to the ratio of the amplitude of the positive pulse to the amplitude of the negative pulse in the waveform (the greater the ratio deviates from 1, the stronger the asymmetry).

[0144] Autocorrelation decay refers to the rate at which the similarity of signals at different time points decreases as the time interval increases (e.g., the similarity of signals at 1-second intervals is 0.8, and at 5-second intervals it is 0.3). The structural stability index is calculated based on the decay rate; the faster the decay, the more unstable the rock mass (the faster the internal fractures develop).

[0145] The weight coefficients of each channel are dynamically allocated based on the tunnel surrounding rock type and real-time construction progress information to form a comprehensive evaluation value; among them, the weight coefficient of the first frequency band is not less than 0.5;

[0146] In some embodiments, the "weighting coefficient" of each channel, namely the processing channels of the first, second, and third frequency bands, reflects the degree of contribution of each channel to the comprehensive evaluation value, and needs to be dynamically adjusted in combination with the surrounding rock type (hard rock / soft rock) and construction progress (such as tunnel face advancement and support operations).

[0147] For example, if the current rock is Class V soft rock (easily fractured) and in the tunneling stage (large disturbance of surrounding rock), the weight of the first frequency band (core frequency band) is set to 0.6, the weight of the second frequency band to 0.2, and the weight of the third frequency band to 0.2 (highlighting low-frequency precursors); if the rock is Class II hard rock and in the stabilization stage after support, the weight of the first frequency band is set to 0.5, and the weight of the second and third frequency bands to 0.25 each (more balanced multi-frequency characteristics of hard rock fracture precursors). The comprehensive evaluation value = (first channel result × 0.6) + (second channel result × 0.2) + (third channel result × 0.2).

[0148] The dynamic feature learning module optimizes the weight coefficients and adaptive thresholds of each channel. The dynamic feature learning module uses historical strain data and early warning records to train the neural network model. The adaptive threshold increases with the tunnel burial depth.

[0149] In practice, historical data of similar tunnels from the past three years are collected: the input is "rock microstrain data (divided into three frequency bands)", and the output is "whether a rupture has actually occurred (1=yes, 0=no)"; a training set (80% of the data) and a validation set (20%) are constructed and trained using a CNN neural network: the input layer receives the three frequency band data, the hidden layer extracts spatiotemporal features, and the output layer predicts "whether a rupture has occurred"; during training, network parameters (such as the size of the convolution kernel) are adjusted until the prediction accuracy of the validation set is ≥90%, ensuring that the model can accurately identify new rupture precursors.

[0150] When the comprehensive evaluation value exceeds the adaptive threshold, pre-trigger verification is performed, including: combining real-time construction progress information to simulate the propagation path of low-frequency strain fluctuations and verify whether it conforms to the spatiotemporal evolution pattern of rock mass fracture precursors.

[0151] Among them, construction progress information includes the current location of the tunnel face (e.g., K1+200) and the length of the supported section; propagation path simulation refers to calculating the diffusion path and timing of low-frequency strain fluctuations from the initial monitoring point to the surrounding sensing units; and spatiotemporal evolution model refers to the propagation law of rock mass fluctuations before fracturing in historical data (e.g., uniform diffusion from the fracture center to both ends).

[0152] For example, the construction progress shows that the tunneling has reached K1+200, and a certain sensing unit (located at K1+200) detects low-frequency fluctuations; the simulated propagation path is: the fluctuations should spread from K1+200 to K1+190 (upstream) and K1+210 (downstream) at a speed of about 0.5 m / s (which is consistent with the propagation speed of historical rupture precursors); if similar fluctuations are actually detected at K1+190 after 10 seconds and at K1+210 after 10 seconds, then it conforms to the evolution pattern and passes the verification.

[0153] In one implementation, the comprehensive evaluation value is determined based on the following formula:

[0154] in, This is a comprehensive assessment value for precursors of rock mass fracture, used to compare with an adaptive threshold to determine whether a stress warning has been triggered. The weighting coefficient for the first frequency band (0.1Hz-1Hz) is 0.5-0.7 (0.7 for hard rock and 0.5 for soft rock). The weighting coefficient for the second frequency band (1Hz-10Hz) is 0.2-0.3, decreasing by 0.05 for every 100m increase in burial depth. The weighting coefficient for the third frequency band (full band) is 0.1-0.2, with 0.2 for fault regions. The normalized value of the energy characteristics of the first frequency band is (energy integral within the time window ÷ historical maximum safe value, range 0-1). The pulse asymmetry index for the second frequency band is calculated as (positive pulse amplitude ÷ negative pulse amplitude, ranging from 0 to 1). The third frequency band structural stability index is (autocorrelation attenuation rate ÷ reference rate, range 0-1). This is a dynamic feature learning correction factor (neural network output, range -0.2 to 0.3).

[0155] After pre-triggered verification, graded stress warning signals are generated;

[0156] In practice, stress warning signals are divided into multiple levels (such as levels I-III) based on the degree to which the comprehensive assessment value exceeds the adaptive threshold. The higher the level, the more urgent the risk of rock mass fracture.

[0157] Examples of grading:

[0158] Level I Warning: The comprehensive assessment value exceeds the threshold by 10%-30% (minor risk), and the signal is marked as "Caution";

[0159] Level II Warning: Exceeding 30%-50% (moderate risk), the signal is marked as "Warning";

[0160] Level III Warning: If the risk exceeds 50% (high risk), the signal is marked as "emergency".

[0161] Different levels correspond to different levels of subsequent ventilation control intensity (e.g., Level III corresponds to the highest ventilation power).

[0162] After generating a stress warning signal, spatial consistency verification is performed, including: taking the sensor unit that triggered the warning as the central reference; acquiring monitoring data of all adjacent sensor units within 50 meters upstream and 50 meters downstream of the central reference; calculating the similarity of the following three indicators between the central reference and each adjacent sensor unit: consistency of the trend of energy accumulation slope, synchronicity of the fluctuation of the pulse asymmetry index, and correlation of the evolution of the structural instability coefficient; when the comprehensive score of the similarity of the three indicators of more than 50% of the adjacent units is lower than the first score threshold, manual on-site verification is triggered.

[0163] Among them, taking the sensing unit that triggers the early warning as the center, the consistency between its monitoring data and that of the surrounding units is evaluated to determine whether the early warning is due to local interference (such as loose fiber optic cables) or overall rock mass change.

[0164] The consistency of the energy accumulation slope is used to indicate whether the energy accumulation slopes of the central unit and adjacent units (e.g., an increase of 5 με per 10 seconds) increase or decrease together (e.g., if the central slope is +5 με / 10s, the adjacent unit is +4 με / 10s, which is consistent; if it is -2 με / 10s, they are not consistent). The synchronicity of the fluctuation of the pulse asymmetry index is used to indicate whether the exponential change trends of the two are synchronized (e.g., if the central index rises from 1 to 2.5, the adjacent unit rises from 1 to 2.3 simultaneously, which is synchronized; if it falls to 0.8, they are not synchronized). The correlation of the evolution of the structural instability coefficient is used to indicate whether the coefficients of the two (e.g., from 0.8 to 0.3) change in the same direction (e.g., if both the central and adjacent units decrease, they are correlated; if one increases and the other decreases, they are not correlated).

[0165] The first scoring threshold is a critical value (range 0-1, higher values ​​indicate stricter similarity requirements) used to judge whether the similarity of the changing trends of the three indicators meets the standard. Values ​​below this threshold are considered dissimilar. For example, in hard rock areas, the first scoring threshold is set at 0.6 (hard rock structures are stable, and high consistency of fluctuations is required); in fault fracture zones, it is set at 0.5 (the rock mass itself is not uniform, allowing for slightly lower consistency). If the comprehensive score of more than 50% of adjacent units is lower than this threshold (e.g., lower than 0.6 in hard rock), manual review is triggered to avoid misjudgment due to local interference.

[0166] Based on the embodiments provided in this application, corrosion-resistant armored optical fibers, combined with sensing units spaced 10 meters apart, not only meet the adaptability requirements of the complex terrain of the tunnel's initial support surface but also achieve high-density, high-accuracy monitoring of rock mass micro-strain. The serpentine laying method further enhances the sensitivity of capturing rock mass deformation. Multi-band processing decomposes the data into 0.1Hz-1Hz, 1Hz-10Hz, and the full frequency band, specifically focusing on the low-frequency core features of precursors to rock mass fracture. The first frequency band has a weight allocation of no less than 0.5, ensuring priority identification of key signals. The dynamic feature learning module trains the neural network using historical data, enabling the weight coefficients and adaptive thresholds to be dynamically adjusted with the tunnel's burial depth, breaking the limitations of traditional fixed threshold early warning. Pre-trigger verification, combined with construction progress simulation of the fluctuation propagation path, and spatial consistency verification, through the similarity assessment of three indicators of adjacent units within a 50-meter range, fundamentally reduces the probability of false triggering by a single sensing unit from a technical perspective. This upgrades stress early warning from a simple signal response to an intelligent judgment with spatial correlation and historical adaptability, providing a more valuable triggering basis for subsequent ventilation control.

[0167] Furthermore, such as Figure 4 As shown, the strain energy-gas release correlation model is constructed based on the following steps:

[0168] S401, Collect rock debris samples generated at the tunnel face in the tunnel construction area;

[0169] S402, mineral composition analysis was performed on the rock slag sample to establish a three-dimensional mineral distribution model; each spatial unit in the three-dimensional mineral distribution model includes the gradient of silica-calcium ratio, the angle of the main fracture direction, and crystallinity parameters.

[0170] In practice, rock debris samples generated from the blasting at the tunnel face are collected (one sample is collected every 5 meters of tunneling, each sample weighing approximately 5 kg). The content of elements such as silicon (Si) and calcium (Ca) is analyzed using an X-ray diffractometer (XRD) to determine the silicon-to-calcium ratio. The direction of fractures and crystallinity (the regularity of crystal arrangement) are observed using a scanning electron microscope (SEM). The analytical data are correlated with the tunneling location (e.g., station number K1+200), and the discrete point data is transformed into a three-dimensional spatial distribution (within a 50-meter range centered on the tunnel face) using a Kriging interpolation algorithm. Each spatial unit (voxel) contains the gradient of silicon-to-calcium ratio change (e.g., an increase of 0.2 per meter), the angle of the main fracture direction (e.g., 30°), and crystallinity parameters (e.g., 0.8, with 1 being the most regular).

[0171] S403, construct a three-dimensional topological network, and extract features by scanning the three-dimensional mineral distribution model through a sliding window. The window size is dynamically adjusted to a 3×3×3 or 5×3×1 voxel structure according to the principal stress direction of the surrounding rock.

[0172] The 3D topological network is a spatial relationship network constructed based on a 3D mineral distribution model, using nodes (representing mineral grains or fracture intersections) and edges (representing connections between grains or fracture extension directions). It is used to capture the spatial connection patterns of minerals and fractures (e.g., "a certain fracture is connected to 3 secondary fractures") and reflect the correlation of the internal structure of the rock mass (e.g., the denser the fracture network, the more gas release channels). The window is a 3D cube (voxel structure). Sliding along the 3D topological network extracts local features such as mineral distribution and fracture connections within the window. For example, when the window slides to a certain area, the combined feature of "silicon-to-calcium ratio gradient 0.3 / m + 3 intersecting fractures + crystallinity 0.6" is extracted as the "fingerprint" of the geological characteristics of that area.

[0173] The direction of the principal stress in the surrounding rock (such as along the tunnel axis or perpendicular to the axis) determines the main development direction of the fractures, and the window size needs to match the fracture direction to accurately capture the features.

[0174] In practical implementation, if the principal stress is along the tunnel axis (length direction), the cracks mostly extend axially. A 5×3×1 voxel (5 voxels along the axis, 3 horizontally and 1 vertically) is used to accommodate long, narrow cracks. If the principal stress is perpendicular to the axis (e.g., radially), the cracks are distributed in all directions. A 3×3×3 voxel (cube) is used to capture the characteristics in all directions evenly. For example, during shield tunneling, if the principal stress is measured along the axis (K1+200 to K1+300 direction), the window automatically switches to a 5×3×1 voxel, making it easier to capture long cracks extending axially.

[0175] S404, set up a dual-mode parameter library, wherein the first mode parameter library stores the baseline parameters of the complete geological profile, and the second mode parameter library stores the correction parameters of the fault-karst region;

[0176] S405: Calculate the similarity score between the current geological features and the data stored in the first model parameter library in real time; when the similarity score exceeds the second scoring threshold, call the benchmark parameters; when the similarity score does not exceed the second scoring threshold, search for cases within a 20-meter range before and after the current tunneling position in the second model parameter library, select the case with the closest spatial distance to generate compensation parameters;

[0177] The second scoring threshold is used to determine the similarity between the current geological features and the first model parameter library (complete geological profile). If the similarity exceeds the threshold, the baseline parameter is used; otherwise, the compensation parameter is used.

[0178] Among them, the benchmark parameters (intact geology) reflect the normal geological characteristics of a faultless and karst cave, including but not limited to the gradient of the silica-calcium ratio: 0.2 / meter (the silica-calcium ratio increases by 0.2 for every 1 meter of depth); the angle of the main fracture direction: 30° (the angle with the tunnel axis); and the crystallinity parameter: 0.9 (the crystals are relatively well-arranged).

[0179] The correction parameters (fault-cave area) are adjustment values ​​for areas with geological anomalies, including but not limited to the gradient of silica-calcium ratio change: 0.5 / m (the mineral distribution near the fault is uneven, and the gradient is larger); the angle of the main fracture direction: 60° (the fault causes the fracture direction to deflect); and the crystallinity parameter: 0.5 (the crystals are more broken due to the compression of the fault).

[0180] For example, in areas with intact rock strata (such as homogeneous sandstone): the second scoring threshold is set to 0.8 (requiring high similarity, as the geology is stable and the baseline parameters are highly applicable); in areas affected by faults (geologically complex): the threshold is set to 0.6 (allowing lower similarity, as the geology near faults is variable and requires more correction).

[0181] S406, air pressure sensors are installed at 5-meter intervals on the tunnel arch to monitor the temporal relationship between air pressure and rock micro-strain data collected by the strain monitoring network;

[0182] S407, when gas pressure fluctuations are detected to precede rock mass micro-strain, the inversion algorithm type is selected according to the preceding time length;

[0183] In the S408 model training process, the input layer receives historical rock mass micro-strain data and corresponding stress warning signals; the hidden layer processes the input data through spatiotemporal convolution; the physical rule verification module performs three verifications in real time: the predicted gas release kinetic energy does not exceed the upper limit of the proportion of input strain energy; the seepage rate is less than the upper limit of the current rock mass permeability; the gas volume and confining pressure maintain a negative exponential relationship; and the output layer generates gas risk level and ventilation control instructions.

[0184] Strain energy is the elastic potential energy (unit: J / m³) stored in the rock mass under stress, while gas release kinetic energy is the kinetic energy (unit: J / m³) possessed by gas when it is released from rock fissures. According to energy conversion laws, gas release kinetic energy originates from the conversion of rock mass strain energy. However, limited by the rock mass structure (such as the degree of fissure development), the conversion ratio has a maximum value, i.e., the upper limit of the strain energy ratio. This upper limit is used to constrain the prediction results of the correlation model, avoiding unreasonable outputs where gas release kinetic energy exceeds the strain energy conversion limit. For example, for Class III intact surrounding rock (few fissures, low energy conversion efficiency): the upper limit is set at 30% (i.e., gas release kinetic energy ≤ strain energy × 30%). For example, when the rock mass strain energy is 1000 J / m³, the model-predicted gas release kinetic energy should not exceed 300 J / m³; for Class V fractured surrounding rock (well-developed fissures, high energy conversion efficiency): the upper limit is set at 50%. For example, when the strain energy is 800 J / m³, the kinetic energy released by the gas must not exceed 400 J / m³.

[0185] Permeability is a parameter (unit: m²) that measures the ability of a rock mass to allow fluids such as gas to pass through, reflecting the conductivity of fracture channels. The current rock mass permeability upper limit refers to the maximum permeability value that the rock mass can achieve under current geological conditions (such as surrounding rock type, burial depth, and fracture density). The physical rule verification requires that "the seepage rate < the rate corresponding to the current rock mass permeability upper limit" to ensure that the gas seepage velocity predicted by the model does not exceed the actual maximum allowable flow capacity of the rock mass, conforming to the laws of fluid mechanics. For example, for sandstone (hard, sparsely fractured): the current rock mass permeability upper limit is set to 1 × 10⁻⁶ m². -15 m² (corresponding to an upper limit of seepage rate of approximately 0.05 m / s). If the model predicts a gas seepage rate of 0.06 m / s in a certain area, the verification will fail, and the model parameters need to be corrected; Mudstone interbedded with coal (weak, densely fractured): The current upper limit of rock mass permeability is set to 5 × 10 m². -14 m² (corresponding to an upper limit of seepage rate of approximately 0.3 m / s). If the predicted seepage rate is 0.25 m / s, it meets the upper limit requirement, and the verification passes. Results are automatically discarded and recalculated when the output violates physical rules.

[0186] Based on the embodiments provided in this application, a three-dimensional mineral distribution model of rock debris samples is used to quantify microscopic features such as the gradient of silica-calcium ratio variation and the direction of the main fracture into model inputs, providing a visual basis for the correlation between the intrinsic properties of the rock mass and gas release. A dual-mode parameter library distinguishes between complete geological profiles and fault-cave areas, and combined with case searches within a 20-meter range, it solves the problem of insufficient parameter adaptability under complex geological conditions. The monitoring and inversion algorithm selection for the relationship between gas pressure and strain time series reveals the dynamic coupling law between gas release and rock mass deformation, while the physical rule verification module constrains the relationship between gas release kinetic energy, seepage rate, gas volume, and confining pressure, ensuring that the model output does not deviate from the engineering physics essence. This design, which combines microscopic geological features, dynamic physical processes, and model constraints, transforms the correlation model from a simple data fitting tool into an intelligent analysis system that reflects the intrinsic mechanism of gas release. The output gas risk level and ventilation control instructions are more closely aligned with the actual situation of gas-rock mass interaction during tunnel construction, upgrading ventilation control from a passive response to an active prediction based on geological mechanisms.

[0187] Furthermore, the operation of the dual-mode parameter library includes:

[0188] Establish a spatial coordinate index system based on the tunnel design centerline stationing, and divide the tunnel axis into several storage sections;

[0189] When the similarity score does not exceed the second score threshold, the spatial location retrieval process is initiated, including: determining the station position of the current tunnel face; and retrieving the adjacent segment case with the smallest distance from the station position of the current tunnel face in the second mode parameter library.

[0190] The differences between the case characteristics of adjacent sections and the current geological characteristics are calculated, including: taking the absolute value of the difference in the silica-calcium ratio gradient; multiplying the difference in the fracture direction angle by a weighting coefficient; taking the absolute value of the difference in the crystallinity parameter; and summing the three difference values ​​to obtain the total difference value.

[0191] Among them, the absolute value of the silicon-calcium ratio gradient difference directly reflects the difference in mineral composition distribution; the fracture direction has a greater impact on gas seepage, so it is multiplied by a weight; the absolute value of the crystallinity parameter difference reflects the difference in crystal integrity; when the total difference value exceeds the set difference threshold, parameter generation is performed, including: analyzing the degree of silicon-calcium ratio anomaly in the analytical data of rock slag samples; assessing the deviation between the current area fracture density and the density benchmark value; calculating the parameter compensation base based on the tunnel burial depth, and applying the Kalman filter algorithm to smooth and generate compensation parameters;

[0192] The difference threshold is set to determine whether the differences between the case features and the current geological conditions are too large. If the differences are too large, new compensation parameters need to be generated. For example, in gently dipping rock strata areas, the difference threshold is set to 8 (case parameters can be reused when the differences are small); in fractured zones, the difference threshold is set to 12 (larger differences are allowed due to the more complex geology).

[0193] The density benchmark value is the standard value of fracture density for a certain type of surrounding rock (unit: fractures / meter, the number of fractures per meter), used to assess the density of fractures in the current area. For example, Class III surrounding rock (relatively intact): density benchmark value = 2 fractures / meter; Class V surrounding rock (fractured): density benchmark value = 5 fractures / meter.

[0194] The parameter fusion process includes: mapping the similarity score to the baseline parameter weight; setting the compensation parameter weight to 1 minus the baseline parameter weight; and fusing the baseline parameter and compensation parameter into the output parameter according to their respective weights.

[0195] In one implementation, the dual-mode parameter fusion output formula is:

[0196] in, The density of rock fractures after fusion (fractures / m³) 3 The number of fractures per unit volume of rock mass is used as input to the gas release model. Reference fracture density (fractures / m) 3 The average value of a complete geological profile, such as 15 sections / m for Class IV surrounding rock. 3 ); To compensate for the fracture density (fractures / m) 3 Correction value for fault zones, such as 35 faults / m in fault zones. 3 ); Geological similarity score (the degree of matching between the current geology and the benchmark profile, ranging from 0 to 1); Correction amount for geological differences (strips / m)3 The compensation value when the total difference exceeds the threshold, 0-5 items / m 3 ); The distance to the nearest fault (m, calculated by mileage marker); , where m is the spatial attenuation coefficient (10m for strong permeability and 5m for weak permeability, controlling the effect of faults on the attenuation rate).

[0197] The dual-mode parameter library is updated every 10 meters of tunneling progress.

[0198] It should be explained that the tunnel advance is the actual length of the tunnel excavation (along the axial direction), and every 10 meters of advance means every 10 meters of excavation (e.g., from station K1+200 to K1+210). Based on the embodiments provided in this application, a full lifecycle management system for parameters from retrieval, generation to updating is constructed. Using the tunnel centerline station as a reference for spatial coordinate indexing, the axial direction is divided into storage segments, enabling parameter retrieval to accurately locate the geological corresponding area of ​​the current excavation face, solving the problems of fuzzy retrieval and low matching efficiency in traditional parameter databases. The differential processing of differences in silica-calcium ratio gradient difference, fracture direction angle difference, and crystallinity parameter difference in the total difference value calculation achieves a quantitative assessment of geological characteristic differences, providing a scientific basis for the generation of compensation parameters. The Kalman filter algorithm smooths the compensation parameters, combined with a weighted fusion mechanism, balancing the stability of the baseline parameters and the immediacy of the compensation parameters. The parameter database is updated every 10 meters of tunnel advance, allowing the model to absorb new geological information in real time. This design upgrades the dual-mode parameter library from a static set of parameters to a dynamic system with self-updating capabilities and accurate spatial matching. This ensures that the associated model remains highly adapted to the current geological conditions as tunnel construction progresses, guaranteeing the continuous accuracy of gas risk assessment from a parameter perspective.

[0199] Furthermore, the execution of the inversion algorithm (performing a complete inversion calculation every 5 minutes) includes:

[0200] When the lead time is between 3 and 5 seconds, the microcrack treatment process is executed.

[0201] When the lead time is between 6 and 8 seconds, execute the macro-fracture handling procedure;

[0202] The updated desorption coefficient value is transmitted to the ventilation control system via an industrial real-time communication protocol.

[0203] In actual implementation, the Profinet protocol is used to send the desorption coefficient from the correlation model server to the fan frequency converter in real time, ensuring timely response of ventilation control.

[0204] The fracture development rate and desorption coefficient are displayed in real time on the monitoring system screen.

[0205] When the desorption coefficient value changes by more than 15% within 1 minute, the audible and visual alarm device is triggered.

[0206] The microcrack treatment process includes:

[0207] The time difference between the start of air pressure fluctuation and the start of rock mass micro-strain was measured using a timer.

[0208] Based on the time difference, the basic rate of fracture development is calculated according to a predetermined linear ratio;

[0209] Among them, in micro-fractures, the time difference between the gas pressure fluctuation and the micro-strain of the rock mass is directly proportional to the fracture development rate (the larger the time difference, the faster the fracture propagates). For example, if the predetermined linear ratio is "rate = 0.17 × time difference" (unit: mm / s), if the time difference is 3 seconds, the rate = 0.17 × 3 = 0.51 mm / s; if the time difference is 5 seconds, the rate = 0.17 × 5 = 0.85 mm / s (which conforms to the law of gradual expansion of micro-fractures over time).

[0210] The silicon-to-calcium ratio was obtained from the current analysis results of the rock debris at the tunnel face;

[0211] Compare the silicon-calcium ratio value with the baseline value of 2.5; when the silicon-calcium ratio value is greater than 2.5, increase the correction amount by 0.02 to the basic fracture development rate value; when the silicon-calcium ratio value is less than 1.8, decrease the correction amount by 0.01 to the basic fracture development rate value.

[0212] It should be explained that the silica-calcium ratio reflects the brittleness of the rock mass (high silica content makes it brittle, while high calcium content makes it tough). A silica-calcium ratio > 2.5 (e.g., 3.0): the rock mass is relatively brittle, and cracks tend to propagate rapidly, thus increasing the crack development rate (+0.02 mm / s); a silica-calcium ratio < 1.8 (e.g., 1.5): the rock mass is relatively tough, and cracks propagate more slowly, thus decreasing the rate (-0.01 mm / s).

[0213] Calculate the desorption coefficient adjustment value based on the final fracture development rate value;

[0214] When the calculated desorption coefficient adjustment value exceeds 25% of the first original desorption coefficient, the desorption coefficient adjustment value is limited to 25% of the first original desorption coefficient.

[0215] When the calculated desorption coefficient adjustment value is lower than -25% of the first original desorption coefficient, the desorption coefficient adjustment value is limited to -25% of the first original desorption coefficient.

[0216] The final determined desorption coefficient values ​​are updated in the associated model parameter register.

[0217] In some embodiments, the desorption coefficient adjustment value = final fracture development rate × correction coefficient (e.g., 0.05). For example, if the final fracture development rate is 0.85 mm / s, the adjustment value = 0.85 × 0.05 = 0.0425.

[0218] The first initial desorption coefficient is an initial value set based on historical data (reflecting the inherent ability of the rock mass to release gas). For example, in sandstone areas: the first initial desorption coefficient is set to 0.05 m³ / (kg・d) (0.05 cubic meters of gas are released per kilogram of rock mass per day); in limestone areas: the first initial desorption coefficient is set to 0.03 m³ / (kg・d) (limestone has poor permeability and a low desorption coefficient).

[0219] The associated model parameter register is a software storage unit that stores real-time model parameters for direct access during model calculations. After the desorption coefficient is updated, it is stored in the register. When the model output layer generates ventilation commands, it can directly read the latest value, ensuring that the commands are based on the current parameters.

[0220] The final desorption coefficient value = the first original desorption coefficient + the desorption coefficient adjustment value.

[0221] Based on the embodiments provided in this application, an accurate parameter calibration mechanism is established to address the unique characteristics of gas release under micro-fractures. The linear proportional relationship between time difference and fracture development rate, combined with corrections for the silica-calcium ratio (increased by 0.02 when greater than 2.5, decreased by 0.01 when less than 1.8), quantifies the intrinsic correlation between rock mass mineral composition and fracture development into calculable parameter adjustments. This ensures that the calculation of fracture development rate not only depends on the time dimension but also incorporates the influence of the rock mass's own properties. A ±25% limitation on the desorption coefficient adjustment value avoids drastic parameter changes caused by fluctuations in micro-fracture monitoring data, ensuring the stability of the model input. This design overcomes the limitations of traditional qualitative descriptions of fracture development. Through the coupled analysis of mineral composition and time difference, it achieves accurate control of key parameters for gas release under micro-fractures, enabling the correlation model to more subtly capture the characteristics of gas release at the micro-fracture stage, providing parameter support for ventilation control that is more closely aligned with micro-geological changes.

[0222] Furthermore, the macroscopic fracturing treatment process includes:

[0223] The time difference between the start of air pressure fluctuation and the start of rock mass micro-strain was measured using a timer.

[0224] Based on the time difference, the fracture development rate value is calculated according to a predetermined inverse relationship;

[0225] It should be explained that in macroscopic rupture, the greater the time difference leading the pressure fluctuation, the lower the rupture development rate (because macroscopic rupture often occurs instantaneously, a large time difference indicates that the rupture has tended to stabilize). For example, the predetermined inverse relationship is "rate = 4.8 / time difference" (unit: mm / s). With a time difference of 6 seconds, the rate = 4.8 / 6 = 0.8 mm / s; with a time difference of 8 seconds, the rate = 4.8 / 8 = 0.6 mm / s.

[0226] Obtain the current tunnel depth data from the construction design drawings; calculate the time difference compensation value according to the rule of increasing by 0.3 seconds for every 100 meters of burial depth;

[0227] The fracture development rate value was recalculated using the compensated time difference.

[0228] For example, for every 100-meter increase in burial depth, the time difference compensation value increases by 0.3 seconds (the greater the burial depth, the greater the rock mass stress, and the time difference needs to be corrected).

[0229] Calculate the desorption coefficient adjustment value based on the recalculated fracture development rate value;

[0230] When the calculated desorption coefficient adjustment value exceeds 30% of the second original desorption coefficient, the desorption coefficient adjustment value is limited to 30% of the second original desorption coefficient.

[0231] When the calculated adjustment value is lower than -30% of the second original desorption coefficient, the adjustment value of the desorption coefficient is limited to -30% of the second original desorption coefficient.

[0232] The final determined desorption coefficient values ​​are updated in the associated model parameter register.

[0233] Wherein, the adjustment value = compensated rate × correction coefficient (e.g., 0.06), for example, if the rate is 0.696 mm / s, the adjustment value = 0.696 × 0.06 ≈ 0.0418.

[0234] The second initial desorption coefficient is the initial value for macroscopic fracture areas (usually higher than microscopic fractures, since macroscopic fractures are more likely to release gas), for example, near faults: 0.1 m³ / (kg・d); in karst caves: 0.12 m³ / (kg・d).

[0235] The final desorption coefficient value = the second original desorption coefficient + the desorption coefficient adjustment value.

[0236] Based on the embodiments provided in this application, focusing on the intensity and complexity of gas release during macroscopic fracturing, a parameter adjustment system considering the influence of burial depth is constructed. The inverse relationship between time difference and fracture development rate aligns with the rapid response characteristics of rock mass deformation and gas release during macroscopic fracturing. Adding a 0.3-second time difference compensation value for every 100 meters of burial depth incorporates the key environmental factor of tunnel burial depth into the parameter calculation, making the assessment of fracture development rate more consistent with the mechanical properties of rock mass at different burial depths. The ±30% limitation on the desorption coefficient adjustment value provides a more reasonable buffer zone to address the tendency for significant parameter fluctuations during macroscopic fracturing, ensuring the safety of parameter adjustment. This design combines the dynamic process of macroscopic fracturing with environmental factors, enabling the correlation model to more accurately calculate gas release-related parameters when facing large-scale rock mass fracturing, thus improving the adaptability and reliability of ventilation control in sudden macroscopic fracturing scenarios.

[0237] Furthermore, based on real-time construction progress information, gas risk levels, and ventilation energy efficiency indicators, the fan control parameters and regional power distribution ratios are dynamically adjusted, including:

[0238] An axially discretized model is constructed within the tunnel construction area. Based on the location of the main ventilation fan, the branch nodes of the ventilation duct, and the real-time location of the excavation face, the tunnel construction area is divided into multiple power control units. A baseline power weight is initialized for each power control unit. The baseline power weight is calculated by fusing the following parameters: the equivalent length of the ventilation duct path between the power control unit and the main ventilation fan, the moving average attenuation rate of the three most recent gas concentration samples within the power control unit, and the permeability index of the surrounding rock type corresponding to the power control unit.

[0239] Real-time acquisition of wind speed, wind pressure and gas concentration monitoring data of each power control unit; mapping of gas risk level to global minimum air volume constraint value; conversion of ventilation energy efficiency index into system energy efficiency optimization weight; generation of pre-ventilation intensity coefficient for unexcavated area (such as the 20-meter unexcavated section ahead) based on the tunneling speed deviation rate of real-time construction progress information;

[0240] Among them, there is a corresponding relationship between the gas risk level (such as level I-III) and the minimum ventilation volume to ensure that the gas concentration does not exceed the standard. The mapping is that the level is converted into a specific air volume value through preset rules, which serves as the bottom line requirement of the ventilation system.

[0241] Ventilation energy efficiency indicators (metabolic output / energy input, such as 1.2) reflect the current energy efficiency level of the ventilation system. The conversion is to transform them into weight values ​​(0-1) through a functional relationship. The higher the weight, the more priority should be given to optimizing energy efficiency.

[0242] The tunneling speed deviation rate = (actual tunneling speed - planned speed) / planned speed (e.g., +20% indicates ahead of schedule, -10% indicates behind schedule); the pre-ventilation intensity coefficient (0.8-1.2) is used to adjust the ventilation intensity of the unexcavated area (a higher coefficient means stronger ventilation) to adapt to changes in the tunneling progress in advance.

[0243] A time fractional differential model is established to dynamically calculate the time fractional order based on the autocorrelation characteristics of the rock mass microstrain data collected by the strain monitoring network; a spatial fractional gradient model is established to calculate the spatial fractional order based on the development direction of the surrounding rock fractures; and the dual-order outputs are fused to generate a state prediction sequence for each power control unit at a future preset time (e.g., 5 minutes).

[0244] In some embodiments, a time-fractional differential model is used to describe the variation of rock mass microstrain over time (the fractional order reflects the "memory" of the variation; for example, an order of 0.6 indicates that the current strain is influenced by nearly 60% of historical data). Based on the autocorrelation characteristics of the microstrain data (e.g., a correlation of 0.7 between adjacent 10-minute data points), the order is obtained by fitting using the least squares method (higher correlation results in a higher order, such as 0.8; lower correlation results in a lower order, such as 0.3).

[0245] The spatial fractional gradient model is used to describe the spatial distribution of strain (the fractional order reflects the "non-uniformity" of spatial diffusion). Based on the development direction of fractures in the surrounding rock (e.g., densely packed along the 45° direction), the fractal dimension of the fracture distribution (e.g., 1.2) is calculated and mapped to the spatial order (a higher fractal dimension corresponds to a higher order, e.g., 0.7).

[0246] The dual-order generation state prediction sequence is fused: the time order (e.g., 0.6) and spatial order (e.g., 0.7) are input into the fusion model, and the prediction sequence for the next 30 minutes (e.g., the gas concentration and wind speed change trends of each power control unit) is generated by weighted summation (time weight 0.6 + spatial weight 0.4).

[0247] In one specific implementation, the expression for the power regulation unit state prediction is:

[0248] in, For the future Predicted state values ​​of the power control unit at any given time (unit: m) 3 / min represents the airflow status of this unit). A future time predicted based on the current time; The time fractional order coefficient is 1.2 when the autocorrelation is strong and 0.8 when it is weak. For the first Historical state values ​​at time (unit: m) 3 / min refers to the historical air volume data of this unit, which is used as the input of the time fractional order model; This represents the i-th historical moment before the current moment t, such as i=1 corresponding to t−1 (1 minute before the current moment), i=2 corresponding to t−2 (2 minutes before the current moment), and is the input of the time fractional model. The baseline time (in minutes, fixed at 30 minutes, used for time dimension normalization); For time fractional order (0 < <1, take 0.7 when autocorrelation decays slowly, and 0.3 when it decays quickly). The spatial fractional order coefficient is 0.9 when the fracture development direction is consistent with the ventilation direction, and 0.5 when it is perpendicular to the direction. For the first Current state values ​​of adjacent control units (unit: m) 3 / min refers to the current airflow data of adjacent units, which is used as input to the spatial fractional-order model. For the current unit and the first Spatial distance between adjacent units (unit: m); The baseline distance (m, fixed at 100m, used for spatial dimension normalization); For spatial fractional order (0 < <1, take 0.6 when the fracture density is high, and take 0.2 when the fracture density is low). Error correction term (unit: m) 3 / min, based on Kalman filtering dynamic compensation, range -5 to 5m 3 / min).

[0249] A dual objective function is constructed based on the state prediction sequence. The first objective is to maximize the system energy efficiency weight, and the second objective is to minimize the concentration difference between power control units with different gas risk levels. Three types of physical constraints are embedded: the minimum wind speed threshold of the unit, the upper limit of the total power of the wind turbine cluster, and the allowable fluctuation range of the static pressure difference between adjacent units. The decomposition and coordination algorithm is used to solve for the wind turbine frequency setpoint and power allocation ratio.

[0250] For example, the minimum wind speed threshold for the unit is: tunnel face area (gas source): ≥1.5m / s (to ensure that gas does not accumulate); rear support area (personnel work area): ≥0.5m / s (to meet personnel breathing requirements).

[0251] Maximum total power of wind turbine cluster: Daily construction: ≤500kW (balanced energy consumption); Emergency warning (Level III risk): ≤800kW (short-term full-load operation is allowed).

[0252] Allowable fluctuation range of static pressure difference between adjacent units: Horizontal tunnel section: ±50Pa (pressure difference is stable to avoid airflow turbulence); Inclined shaft section with slope >10°: ±80Pa (slope causes pressure difference to fluctuate easily, so the range is relaxed).

[0253] The decomposition and coordination algorithm is used to decompose the optimization problem of multiple power control units into "sub-problems (single unit)" and "coordination problems (global)," and solve the wind turbine parameters step by step. In some embodiments, the solution steps of the decomposition and coordination algorithm are as follows:

[0254] Decomposition: The tunnel is divided into 5 power control units, each unit is treated as a sub-problem, with the goal of "maximum energy efficiency of itself + pressure difference with adjacent units meeting the standard", and the fan frequency is initially calculated (e.g., unit 1 = 40Hz, unit 2 = 45Hz).

[0255] Coordination: The global layer checks whether the total power exceeds the upper limit (e.g., the total power of 5 sub-units is 550kW, which exceeds the upper limit of 500kW). By adjusting the weights (reducing the power weight of high-risk units), the frequency of unit 2 is reduced to 42Hz, and the total power is 480kW.

[0256] Iteration: Repeatedly decompose and coordinate until all constraints are met (e.g., total power 480kW, wind speed of each unit ≥ threshold), and output the final frequency setting value and power distribution ratio.

[0257] Based on the embodiments provided in this application, ventilation control has been upgraded from overall extensive regulation to precise regional management. Power control units, divided according to the location of the main fan, duct branch nodes, and tunnel face, enable ventilation control to accurately meet the actual needs of different construction areas in the tunnel. The benchmark power weight integrates multiple parameters such as duct path resistance, gas concentration decay rate, and surrounding rock permeability, providing a multi-dimensional scientific basis for initial power allocation. The fusion of the time fractional-order differential model and the spatial fractional-order gradient model, through the autocorrelation characteristics of rock mass micro-strain and the direction of surrounding rock fractures, allows the state prediction sequence to more accurately reflect future ventilation state changes in different areas. The dual objective function considers both energy efficiency and concentration differences, and the embedded unit wind speed, total power, and static pressure difference constraints ensure the engineering feasibility of the optimization results. The application of the decomposition and coordination algorithm achieves global optimization of multi-unit control parameters. This design transforms the adjustment of fan control parameters and power distribution ratios from rough settings based on experience to intelligent decision-making based on regional characteristics, predictive models, and multi-objective optimization. While ensuring construction safety in each area, it maximizes the overall energy efficiency of the ventilation system.

[0258] Furthermore, when a stress warning signal is received or a gas risk level increase event is detected, the spatial correction coefficient is updated with the warning signal trigger location as the center point, and the control weight of each power control unit within a radius of 50 meters is enhanced; the dynamic time parameters are adjusted according to the main frequency characteristics of rock mass micro-strain fluctuations.

[0259] Among them, the gas risk level upgrade event refers to the situation where the gas risk level rises from a lower level to a higher level, which requires the triggering of enhanced ventilation control.

[0260] Among them, the dominant frequency characteristic of rock mass microstrain fluctuation refers to the frequency with the highest energy in microstrain fluctuation (e.g., 0.5Hz as the dominant frequency); the dynamic time parameter is the time window for predicting the future state (e.g., 5-15 minutes). A higher dominant frequency means faster fluctuation, and the window needs to be shortened to improve the response speed. For example, dominant frequency 0.8Hz (frequent fluctuation): dynamic time parameter = 5 minutes (prediction updated every 5 minutes); dominant frequency 0.2Hz (smooth fluctuation): parameter = 15 minutes (extending the prediction window and reducing the amount of calculation).

[0261] The state prediction sequence is compared with the actual monitoring data every 2 minutes; when the prediction deviation of a certain power control unit exceeds the set allowable threshold, the boundary division of the power control unit is adjusted according to the rock joint orientation of the geological sketch data; the resistance coefficient is updated based on the measured data of the air duct pressure sensor; and the air permeability index is corrected by integrating the current rock slag mineral analysis results of the tunnel face.

[0262] The set allowable thresholds are used to determine whether the deviation between the state prediction sequence and the actual monitoring data is acceptable. If the deviation is exceeded, the model needs to be corrected. For example, the allowable threshold for gas concentration prediction deviation is ±0.1% (if the prediction is 0.5% and the actual deviation is 0.65%, the threshold is exceeded); the allowable threshold for wind speed prediction deviation is ±0.2m / s (if the prediction is 1.5m / s and the actual deviation is 1.2m / s, the threshold is exceeded).

[0263] Geological sketch data records the orientation of rock strata joints (fractures) (e.g., extending along a 30° direction); the boundary is adjusted so that the unit boundary is parallel to the joint orientation, avoiding ventilation control distortion caused by dividing across joints. For example, if the original unit boundary is divided along the tunnel axis (0°), and the geological sketch shows a joint orientation of 30°, the boundary is adjusted to a 30° direction to ensure consistent rock permeability within the same unit and more precise ventilation control.

[0264] The drag coefficient reflects the friction resistance along the duct (e.g., 5 Pa / m). The pressure sensor measures the air pressure loss of a certain section of the duct (e.g., 100 Pa / 20 meters, i.e., 5 Pa / m). If the deviation from the original coefficient (4 Pa / m) exceeds 10%, it is updated to the measured value of 5 Pa / m to ensure accurate air volume calculation.

[0265] The permeability index reflects the rock mass's ability to allow gas infiltration (e.g., 0.8). Rock debris analysis shows an increase in the silica-calcium ratio (the rock mass becomes more brittle and fissures increase), so the permeability index is corrected (e.g., from 0.8 to 1.0) to make the model more closely match the current gas infiltration characteristics of the rock mass.

[0266] Measure the response delay time of each fan control loop; calculate the command pre-compensation offset according to dynamic time parameters; output the frequency setpoint before the planned issuance time.

[0267] The response delay time refers to the lag between the wind turbine receiving the command and the actual speed adjustment (e.g., 2 seconds); the pre-compensation offset = delay time × wind speed change rate (e.g., 0.5 m / s²), which is 1 m / s; the planned "wind speed + 1 m / s" command was issued at 10:00, but it was actually issued at 9:59:58, compensating for the 2-second delay to ensure that the target wind speed is reached on time at 10:00.

[0268] Collect data from the fan speed sensor; when the detected speed fluctuation exceeds the stable threshold, activate the instruction smoothing module; generate an amplitude attenuation parameter smoothing instruction output based on the spatial correction coefficient;

[0269] The stability threshold refers to the maximum allowable fluctuation range of the rotational speed (e.g., ±5% of the rated speed); the command smoothing process reduces fluctuations by lowering the command change rate (e.g., from "+2Hz / second" to "+0.5Hz / second"). For example, main fan (rated speed 1500r / min): ±75r / min (5%); auxiliary fan (rated speed 1000r / min): ±50r / min (5%).

[0270] The smoothed frequency setpoint of the fan is sent to the frequency converter via the industrial real-time network; the optimized power distribution ratio is written into the regional power distribution controller.

[0271] The smoothed frequency setting (e.g., 42Hz) is sent to the frequency converter via an industrial bus (e.g., Modbus) to directly control the motor speed; the power distribution ratio (e.g., 30% for unit 1 and 25% for unit 2) is written into the area controller to adjust the opening of the air valves in each unit to achieve power distribution.

[0272] Based on the embodiments provided in this application, a closed-loop control system with feedback correction and emergency response capabilities is constructed. During an early warning, the control weight of units within a 50-meter radius of the trigger location is increased. Dynamic time parameters are adjusted by combining the dominant frequency of rock mass micro-strain fluctuations, enabling ventilation control to quickly focus on risk areas in emergency scenarios and improving the targeted nature of the response. Prediction deviation comparison every 2 minutes, combined with adjustments to unit boundaries based on rock joint orientation, updates to resistance coefficients, and corrections to the permeability index, achieves real-time calibration of control parameters, resolving the potential discrepancy between model predictions and actual operating conditions. Pre-compensation for fan response delays and smoothing of commands during speed fluctuations avoid oscillations during control command execution, ensuring the stability of the ventilation system. This design extends dynamic adjustment from parameter calculation to the entire execution feedback process, enabling ventilation control to self-correct and adapt to operating condition fluctuations, improving control reliability and continuity, and ensuring the ventilation system maintains optimal operating conditions in complex construction environments.

[0273] Furthermore, energy inputs include fan energy consumption and waste transport volume, while metabolic outputs include personnel thermal comfort and pollutant removal volume.

[0274] Among them, the energy consumption of the wind turbine refers to the total power consumption of the main wind turbine and the auxiliary wind turbine (unit: kWh). For example, if the main wind turbine consumes 200 kWh and the auxiliary wind turbine consumes 50 kWh in a certain hour, the total energy consumption is 250 kWh.

[0275] The volume of excavated soil transported refers to the volume of rock debris transported during construction (unit: m³). The transportation process will raise dust and disturb airflow, indirectly increasing the ventilation demand. If the daily transportation volume is 500 m³, it needs to be included in the energy input assessment.

[0276] Human thermal comfort is comprehensively assessed based on temperature (18-26℃), humidity (40%-60%), and wind speed (0.3-0.8m / s). For example, if the temperature in a certain area is 24℃ and the humidity is 50%, it is judged as "comfortable" and corresponds to a positive contribution to metabolic output.

[0277] Pollutant removal capacity refers to the total amount of pollutants such as gas and dust discharged by the ventilation system (e.g., gas removal capacity of 50 m³ / h, dust removal capacity of 2 kg / h), which directly reflects the core function and effectiveness of ventilation.

[0278] Based on the embodiments provided in this application, the limitations of traditional methods that only measure ventilation efficiency by fan energy consumption are overcome, and a more comprehensive ventilation system efficiency evaluation system is established. The amount of excavated soil transported is included in the energy input, taking into account the energy consumption of operations other than fans during tunnel construction, making the energy input calculation more closely aligned with actual construction conditions. Metabolic outputs include personnel thermal comfort, focusing on the improvement of the working environment for construction workers by the ventilation system, while the amount of pollutant removal is directly related to the core safety objective of ventilation. This multi-dimensional energy efficiency index calculation ensures that the evaluation of ventilation energy efficiency not only focuses on energy consumption but also takes into account construction safety and personnel comfort, providing a more comprehensive decision-making basis for dynamically adjusting fan parameters. Through this more scientific energy efficiency evaluation, the operation of the ventilation system can achieve the most rational allocation of energy while meeting construction safety and personnel needs, promoting the transformation of ventilation control from simple safety assurance to comprehensive management that emphasizes both safety and energy efficiency.

[0279] It should be noted that the embodiments implemented on the side of the tunnel ventilation control system based on artificial intelligence in this application can be referenced with the embodiments implemented on the side of the tunnel ventilation control method based on artificial intelligence, and will not be described in detail here.

[0280] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An artificial intelligence based tunnel ventilation control system characterized in that, include: The design module is used to establish an unsteady computational fluid dynamics model based on the three-dimensional alignment parameters of the tunnel during the tunnel design phase, calculate the turbulent structure under different traffic conditions using the large eddy simulation method, and determine the optimal spatial configuration scheme of the wind turbine group based on the velocity field and pressure field distribution. The construction module is used to deploy multi-type sensor arrays in layers along the tunnel arch and sidewalls during the tunnel construction phase, including a distributed fiber optic temperature sensing system, a miniature air pressure sensor network, and a laser particulate matter monitor, to build a full-section environmental parameter acquisition system. The data acquisition module is used to collect data on the temperature gradient field, pressure distribution field, and pollutant concentration field of each zone in real time through a sensor array during the tunnel operation phase, and simultaneously obtain real-time vehicle trajectory and speed distribution information from the traffic monitoring system. The prediction module is used to establish a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis. It analyzes environmental multi-physics field data and traffic flow characteristics to predict the migration and diffusion patterns of pollutants in different zones in the future. The establishment of a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis includes: processing environmental monitoring time-series data using an empirical mode decomposition algorithm to extract intrinsic mode functions of pollutant concentration fluctuations at different time scales; constructing a nonlinear mapping network between traffic flow parameters and pollutant diffusion rates, and performing multi-step time-series predictions using a long short-term memory recurrent neural network; dynamically correcting the diffusion coefficient tensor in the pollutant transport equation based on temperature stratification effects and the spatial distribution of longitudinal wind speed; discretizing the tunnel space into a structured control volume grid, solving the unsteady convection and diffusion equations using the finite volume method, and predicting the spatiotemporal evolution of pollutants; and fusing numerical prediction results with real-time monitoring data using an adaptive Kalman filter algorithm to update model parameters and state estimates online. The solution module is used to construct a wind turbine collaborative control model that considers airflow organization optimization. The optimization objective is to minimize the total energy consumption of the system, and the constraint is that the pollutant concentration in each zone does not exceed the threshold. The optimal operation strategy is solved by a multi-objective adaptive weight allocation algorithm. The adjustment module is used to dynamically adjust the operating status of the fans based on the optimization results, and prioritize the activation of downstream fan groups in the area with the largest pollutant concentration gradient to form an efficient ventilation corridor. The optimization module is used to activate the intelligent graded smoke exhaust mode when fire characteristic signals are detected. It dynamically optimizes the smoke exhaust path based on real-time simulation of smoke movement and coordinates multiple sets of fans to form a relay-style smoke exhaust airflow organization. 2.A tunnel ventilation control method based on artificial intelligence, characterized in that, include: During the tunnel design phase, an unsteady computational fluid dynamics model is established based on the tunnel's three-dimensional alignment parameters. The turbulent structure under different traffic conditions is calculated using the large eddy simulation method. The optimal spatial configuration scheme of the wind turbine group is determined based on the velocity field and pressure field distribution. During the tunnel construction phase, multiple types of sensor arrays are arranged in layers along the tunnel arch and sidewalls, including a distributed fiber optic temperature sensing system, a miniature air pressure sensor network, and a laser particulate matter monitor, to build a full-section environmental parameter acquisition system. During the tunnel operation phase, sensor arrays are used to collect real-time data on temperature gradient fields, pressure distribution fields, and pollutant concentration fields in each zone, and simultaneously acquire real-time vehicle trajectory and speed distribution information from the traffic monitoring system. A dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis was established. Environmental multi-physics field data and traffic flow characteristics were analyzed to predict the migration and diffusion patterns of pollutants in different zones in the future. The establishment of a dynamic prediction model for ventilation demand based on spatiotemporal correlation analysis includes: processing environmental monitoring time-series data using an empirical mode decomposition algorithm to extract intrinsic mode functions of pollutant concentration fluctuations at different time scales; constructing a nonlinear mapping network between traffic flow parameters and pollutant diffusion rates, and performing multi-step time-series predictions using a long short-term memory recurrent neural network; dynamically correcting the diffusion coefficient tensor in the pollutant transport equation based on temperature stratification effects and the spatial distribution of longitudinal wind speed; discretizing the tunnel space into a structured control volume grid, solving the unsteady convection and diffusion equations using the finite volume method, and predicting the spatiotemporal evolution of pollutants; and fusing numerical prediction results with real-time monitoring data using an adaptive Kalman filter algorithm to update model parameters and state estimates online. A wind turbine collaborative control model considering airflow organization optimization is constructed. The optimization objective is to minimize the total energy consumption of the system, and the constraint is that the pollutant concentration in each zone does not exceed the threshold. The optimal operation strategy is solved by a multi-objective adaptive weight allocation algorithm. The operating status of the fans is dynamically adjusted based on the optimization results, and the downstream fan groups in the area with the largest pollutant concentration gradient are activated first to form an efficient ventilation corridor. When fire characteristic signals are detected, the intelligent graded smoke exhaust mode is activated. Based on real-time simulation of smoke movement, the smoke exhaust path is dynamically optimized, and multiple sets of fans are coordinated to form a relay-style smoke exhaust airflow organization. 3.The tunnel ventilation control method based on artificial intelligence according to claim 2, wherein, The construction of the nonlinear mapping network between traffic flow parameters and pollutant diffusion rates includes: Establish a spatiotemporal graph convolutional network based on an attention mechanism to represent traffic flow data as a spatiotemporal graph structure; Vehicle density, average speed, and vehicle type composition features are embedded in the nodes of the graph, and vehicle following relationships and lane change information are embedded in the edges of the graph. Local and global spatiotemporal features of traffic flow field are extracted through multi-layer spatiotemporal convolution operations; The extracted features are cross-attention calculations performed with pollutant concentration monitoring data in the latent space; The time dynamics are captured using a gated loop unit, and the predicted pollutant concentration distribution is output for multiple future time steps. 4.The AI-based tunnel ventilation control method of claim 2, wherein, The method of using a multi-objective adaptive weight allocation algorithm to solve for the optimal operating strategy includes: The definition includes a multi-objective optimization function that incorporates energy consumption economic indicators, pollutant distribution uniformity indicators, and airflow stability indicators; Based on the real-time environmental characteristics of each partition, the weight coefficient distribution of each indicator in the objective optimization function is dynamically adjusted based on fuzzy inference. An improved non-dominated sorting genetic algorithm is used to solve the high-dimensional Pareto optimal solution set, and an elite retention strategy is introduced to maintain population diversity. Establish a preference decision-making mechanism based on multi-attribute utility theory to select the optimal control scheme that balances energy saving effect and ventilation quality from the non-dominated solution set; The feasible range of key control parameters was determined through global sensitivity analysis. 5.The tunnel ventilation control method based on artificial intelligence according to claim 4, wherein, The distribution of weight coefficients for each indicator in the objective optimization function based on fuzzy inference includes: The pollutant concentration deviation, concentration change trend, and traffic flow density of each zone are used as fuzzy input variables; Design triangular and trapezoidal membership functions to fuzzify the input variables; Establish a knowledge base that includes multiple fuzzy rules and define weight allocation strategies for different working conditions; The Mamdani fuzzy reasoning method is used for rule reasoning to obtain a fuzzy set of output variables; The centroid method is used for defuzzification, and the real-time weight coefficients of each objective function are output. 6.The AI-based tunnel ventilation control method of claim 2, wherein, The formation of a highly efficient ventilation corridor includes: Gaussian mixture model clustering algorithm is used to identify regions of abnormal concentration of pollutants; Calculate the concentration gradient vector field distribution of each zone to determine the main migration direction and diffusion rate of pollutants; The optimal ventilation path is planned based on the principle of vector synthesis, and the fan groups downstream of the path are activated first. The output power distribution of the fan is adjusted in real time according to the spatiotemporal variation characteristics of the concentration gradient to achieve directional transport and removal of pollutants; Establish a dynamic evaluation mechanism for ventilation effectiveness, and verify the effectiveness of the control strategy through the concentration field uniformity index and energy consumption efficiency index. 7.The tunnel ventilation control method based on artificial intelligence according to claim 2, wherein, The activation of the intelligent graded smoke extraction mode includes: Multi-source information fusion technology is used to identify the location and scale characteristics of fires, including temperature anomaly distribution detection, smoke concentration gradient analysis, and video image pattern recognition; Based on the fire dynamics simulation model, the trajectory of smoke movement is predicted, and the spatiotemporal distribution characteristics of smoke concentration in each zone are calculated. Based on flue gas prediction results, smoke emission control zones are dynamically divided, and graded response control strategies are formulated. A distributed model predictive control method is used to optimize the start-stop timing and speed ratio of the exhaust fan; The pressure field coordinated control algorithm prevents flue gas backflow and fresh air short-circuiting. 8.The tunnel ventilation control method based on artificial intelligence according to claim 7, wherein, The optimization of the smoke extraction scheme using a distributed model predictive control method includes: The tunnel smoke exhaust system is decomposed into multiple mutually coupled sub-control systems; A local prediction model is established for each control subsystem to describe its dynamic characteristics and interactions; Design a distributed objective function to coordinate the control objectives of each subsystem; The alternating direction multiplier method is used to solve distributed optimization problems and achieve coordinated control among subsystems.

Images