Tunnel gas concentration prediction and ventilation scheduling method based on big data analysis

Abstract

The invention discloses a tunnel gas concentration prediction and ventilation scheduling method based on big data driving, and belongs to the technical field of tunnel construction safety ventilation. The invention aims to solve the problems of inaccurate gas prediction, lag in ventilation response and over-high energy consumption caused by complex and changeable geological conditions in the prior art. The method comprises the following steps: clustering historical data through unsupervised learning to divide geological prediction units, and training an exclusive gas concentration prediction model for each unit to form a model library; fusing the mechanical working condition and the rock ballast image data in real time, and dynamically identifying the current geological category by means of a geological identification model; matching an identification result with a model library, if the matching is successful, calling a corresponding model, and if the matching is failed, starting emergency ventilation and generating a temporary model through online incremental learning of the most adjacent model; and finally, according to a concentration prediction value output by the model, optimizing and generating a ventilation scheduling instruction by taking optimal ventilation efficiency and lowest energy consumption as targets. The method is used for realizing accurate prediction of gas and intelligent and efficient control of ventilation in tunnel construction.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction safety and ventilation control technology. More specifically, this invention relates to a big data-driven method for predicting tunnel gas concentration and scheduling ventilation. Background Technology

[0002] During tunnel construction, ensuring a safe gas environment inside the tunnel is crucial. Abnormal changes in the concentration of harmful gases, such as methane, hydrogen sulfide, and carbon monoxide, pose a direct threat to the safety of construction workers and the progress of the project. Currently, this field relies on a range of technologies for gas monitoring and ventilation control; however, these methods still face several inherent challenges in practice that urgently need to be addressed.

[0003] The primary problem lies in the insufficient accuracy of gas concentration prediction. Existing prediction methods largely rely on statistical analysis of historical monitoring data or mathematical models based on fixed geological assumptions. However, tunnel engineering is a typical linear underground engineering project, and the geological conditions of the strata it traverses are often highly complex and variable. Single or simple prediction models are insufficient to accurately depict the dynamic laws of gas dispersion and migration under different lithological, structural, and hydrological conditions. When the tunneling face moves from a homogeneous geological section to another significantly different geological section, the original prediction model may quickly become ineffective, leading to a large deviation between the predicted and actual values. The root cause of this deviation is that the model fails to establish an effective and adaptive correlation with the changing geological environment. Attempting to establish a universal prediction model applicable to all geological conditions presents significant technical difficulties because the formation process of geological bodies is extremely complex, and its inherent laws cannot be exhaustively contained by a single model.

[0004] Secondly, it's difficult to balance real-time performance and cost-effectiveness in ventilation scheduling strategies. Current ventilation control largely relies on fixed thresholds; for example, ventilation is activated or increased when sensors detect gas concentrations exceeding safety limits. This approach is reactive and suffers from response lag, failing to intervene in the early stages when harmful gases begin to accumulate but haven't reached dangerous thresholds. On the other hand, to ensure safety, construction sometimes employs a conservative strategy of continuous high-power ventilation. While this reduces risk, it leads to significant energy waste and high operating costs. Achieving on-demand ventilation—that is, adjusting ventilation equipment in advance based on accurate predictions of future gas concentration trends—is the goal of optimized scheduling. However, the core bottleneck in achieving this goal lies precisely in the inaccuracy of the aforementioned prediction process. Without reliable, forward-looking predictions, optimized scheduling lacks an effective basis for decision-making.

[0005] Finally, the emergency response capability for unforeseen geological anomalies is limited. Tunnel construction is an exploratory project; despite preliminary geological surveys, unfavorable geological formations such as faults, karst caves, and gas-rich cavities may suddenly be encountered ahead of the tunnel face. The geological characteristics of these unforeseen areas differ significantly from existing experience, and the gas escape mechanisms are unclear. In such cases, not only do existing predictive models become completely ineffective, but the lack of effective historical data for the anomaly area also makes it impossible to quickly build new predictive models. This creates a critical safety blind spot: a dangerous decision-making vacuum exists between identifying the anomaly and establishing effective countermeasures. How to quickly acquire and utilize limited real-time data in this emergency situation to rapidly develop a relatively reliable predictive capability and guide emergency ventilation remains a long-standing technical challenge. Traditional model training methods require large amounts of data and long periods, which cannot meet the stringent timeliness requirements under such conditions.

[0006] In summary, existing technologies for tunnel gas concentration prediction and ventilation scheduling mainly face problems such as poor adaptability of prediction models due to variable geological conditions, difficulties in real-time optimization of ventilation scheduling due to inaccurate predictions, and insufficient emergency response capabilities in the face of unforeseen geological anomalies. The root of these problems lies in the contradiction between static models and dynamic environments, as well as the contradiction between data utilization efficiency and decision-making timeliness. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0008] Another objective of this invention is to provide a big data-driven method for predicting tunnel gas concentration and scheduling ventilation, which addresses the technical problems of traditional tunnel gas prediction methods being unable to adapt to complex and variable geological conditions, resulting in inaccurate predictions and consequently relying on delayed responses or conservative strategies for ventilation scheduling, failing to achieve a balance between safety and energy efficiency; and lacking rapid and effective emergency prediction and decision-making capabilities when encountering unforeseen geological anomalies.

[0009] To achieve these objectives and other advantages of the present invention, a method for predicting tunnel gas concentration and scheduling ventilation based on big data is provided, comprising the following steps: S1. Based on the multi-source heterogeneous historical data of the tunnel, geological feature clustering is performed using an unsupervised learning algorithm to divide the tunnel into multiple geological prediction units; S2. For each geological prediction unit, extract multi-dimensional features containing temporal correlations, and use an ensemble learning framework to train a dedicated gas concentration prediction model to form a model library. S3. During tunnel excavation, real-time data on the working conditions of construction machinery and rock debris images at the tunnel face are collected and integrated, and the current geological category is dynamically identified through a pre-trained geological identification model. S4. Match the dynamically identified geological categories with the geological prediction units in the model library: If a match is found, the corresponding dedicated gas concentration prediction model is used as the effective prediction model. If the matching fails, it is determined that the area has entered an unforeseen geological anomaly zone. The emergency ventilation plan is then activated, and the online model learning process is started simultaneously. The online learning process is based on the real-time data stream newly generated in the current anomaly zone. Using the prediction model of the most adjacent geological unit as the base model, incremental learning and rapid fine-tuning are performed to generate a temporary prediction model for the anomaly zone as an effective prediction model. S5. Input real-time environmental data into an effective prediction model to obtain gas concentration prediction values ​​for future periods. Based on these prediction values, and with the goals of optimal ventilation efficiency and minimum energy consumption, generate differentiated ventilation scheduling instructions in real time through an optimization algorithm to drive the ventilation equipment in the corresponding area to execute.

[0010] Preferably, in step S1 of the present invention, the geological feature clustering using an unsupervised learning algorithm specifically involves: Based on historical data such as rock strength, groundwater status, fault structure information, and historical hazardous gas concentrations, a geological-gas feature vector is constructed. The Topplitz inverse covariance clustering algorithm, which introduces mandatory link constraints, is used to cluster geological-gas feature vectors. Among these, it is necessary to link constraints based on expert knowledge construction of historical lithology logging data and gas emission patterns, which is used to guide the algorithm to classify samples with similar lithology and similar gas emission patterns into the same geological prediction unit.

[0011] Preferably, in step S4 of the present invention, the online model learning process specifically includes: From the historical data repository, a portion of historical samples are extracted based on the similarity of data distribution with the current nearest geological unit to form a replay sample set; The replay sample set is mixed with the newly generated real-time data stream in the current anomaly region and used together as training data for incremental learning. During incremental learning and rapid fine-tuning, the replay sample set is weighted more heavily than the new sample set to constrain the model parameters from deviating excessively from the original knowledge during fine-tuning.

[0012] Preferably, in step S4 of the present invention, after determining that an unforeseen geological anomaly zone has been entered, the following steps are also performed: Real-time fusion of rock debris image data from the working face, tunneling parameters of construction machinery, and information on the concentration of trace special gases collected by environmental sensors; Based on the fused information, a pre-trained deep learning model is used to calculate the dynamic safety risk score of the current working surface; When the safety risk score exceeds the preset threshold, a high risk is determined, and the level and response measures of the emergency ventilation plan are adjusted accordingly.

[0013] Preferably, in step S5 of the present invention, with optimal ventilation efficiency and minimum energy consumption as multiple objectives, differentiated ventilation scheduling instructions are generated in real time through an optimization algorithm, specifically implemented through the following steps: S5.1 Construct a comprehensive evaluation index system that includes the average wind speed in the tunnel, the wind speed non-uniformity coefficient, the peak concentration of harmful gases, and the total power of the ventilation system; calculate the membership degree value of each individual index based on the fuzzy logic method; use the improved Seagull optimization algorithm to optimize the membership degree value of each index to obtain the equilibrium solution that makes the membership degree value of all indices optimal as a whole, and determine the required air volume of each branch in the ventilation network to meet the predicted gas concentration value based on the equilibrium solution. S5.2. The tunnel ventilation network is abstracted into a graph structure consisting of nodes and branches, where nodes represent tunnel faces, junctions, or ventilation openings, and branches represent tunnel sections. The wind resistance characteristics of each branch are recorded. The air volume and air pressure of all fans in the network are used as optimization variables, the required air volume obtained in step S5.1 is used as a constraint, and the optimization objective is to minimize the total energy consumption of the fans. The ventilation network optimization model is constructed. The deep deterministic strategy gradient reinforcement learning algorithm is used to solve the ventilation network optimization model. Through continuous interaction between the agent and the ventilation network environment, the fan collaborative control strategy that minimizes the total energy consumption is learned, and ventilation scheduling instructions are generated.

[0014] Preferably, in step S1 of the present invention, the fusion process of multi-source heterogeneous historical data further includes a data augmentation step, which specifically includes: A conditional generative adversarial network is used to generate synthetic samples by using geological category labels obtained by clustering through the Toplitz inverse covariance clustering algorithm as conditional input. Calculate the Mahalanobis distance between the synthetic sample and the cluster center of its geological category, compare the Mahalanobis distance with a preset threshold, and retain the synthetic samples whose Mahalanobis distance is not greater than the preset threshold to obtain the filtered synthetic samples. The selected synthetic samples were added to the geological-gas feature vector set to expand the training dataset used to train the dedicated gas concentration prediction model.

[0015] Preferably, in the online model learning process of step S4 of the present invention, the specific process of incremental learning and rapid fine-tuning includes: establishing a dynamically updated memory bank to store representative historical samples; when performing incremental learning, selecting samples from the memory bank through an importance sampling strategy and mixing them with the real-time data stream.

[0016] Preferably, in step S5.2 of the present invention, the solution process of the ventilation network optimization model specifically includes: modeling the fan cooperative control problem as a partially observable Markov decision process, wherein each fan is an agent, and a multi-agent reinforcement learning algorithm is used for distributed training, and local observation information is shared through the communication mechanism between agents.

[0017] Preferably, the calculation process of the dynamic safety risk score of the present invention further includes: constructing a multimodal fusion module, extracting features from rock debris image data, tunneling parameters and trace special gas concentration information through convolutional neural networks, fully connected layers and temporal encoders respectively, and fusing these features using a cross-modal attention mechanism.

[0018] Preferably, in step S5.2 of the present invention, when constructing the ventilation network optimization model, physical constraints are further introduced, and the Darcy-Weisbach equation and continuity equation in fluid mechanics are embedded as constraints into the reward function of the deep deterministic policy gradient reinforcement learning algorithm.

[0019] The present invention has at least the following beneficial effects: 1. This invention achieves precise matching between the prediction model and geological conditions through geological clustering and the construction of a dedicated model library, significantly improving the accuracy and scenario adaptability of gas concentration prediction. Its dynamic geological identification and matching mechanism ensures that the prediction model can adaptively switch during the tunneling process, enhancing the system's robustness. For unforeseen geological anomalies, the linked response of emergency ventilation and online model learning provides both immediate safety assurance and rapid reconstruction of prediction capabilities, effectively shortening the vacuum period for safety decision-making. Finally, based on accurate prediction and multi-objective optimized scheduling, on-demand allocation of ventilation resources can be achieved, fundamentally resolving the contradiction between safety and energy consumption, and realizing intelligent and lean management of tunnel ventilation.

[0020] 2. This invention introduces a mandatory link constraint based on expert knowledge, combining engineering experience with data-driven algorithms. This makes the geological clustering results not only data-cohesive but also possess clear engineering and physical significance. This ensures higher consistency in geological genesis and gas emission behavior within each geological prediction unit, laying a reliable physical foundation for subsequent training of high-precision dedicated gas concentration prediction models, thereby improving the accuracy and reliability of the entire prediction chain from the source.

[0021] 3. This invention effectively mitigates the "catastrophic forgetting" problem in incremental learning by introducing a historical replay sample set and assigning it higher weights. This strategy enables the model to rapidly adapt to subtle changes in features of new geological areas while maintaining its general predictive ability across a wide range of geological types, ensuring the stability and generalization performance of the temporary prediction model. This makes the system not only respond quickly to unknown anomalies but also produce more reliable prediction results.

[0022] 4. This invention achieves comprehensive and proactive perception of the safety status of unforeseen geological anomalies by integrating multi-source information such as rock debris, machinery, and gases for dynamic risk scoring. This transforms emergency response from a passive reaction based solely on gas concentration into proactive decision-making based on multi-dimensional risk assessment. Adjusting ventilation plans according to the scoring level enables precise allocation of emergency resources, avoiding under- or over-response, and significantly improving the safety management level of high-risk geological sections.

[0023] 5. This invention decomposes ventilation scheduling into two closely linked stages: optimized allocation of demand air volume and coordinated optimization of fan energy consumption. The first stage, through fuzzy comprehensive evaluation and improved optimization algorithms, scientifically balances multiple contradictory indicators within ventilation efficiency, determining a reasonable demand air volume. The second stage leverages the advantages of reinforcement learning in handling complex nonlinear optimization problems, directly optimizing the fan operation strategy while satisfying air volume constraints, thus minimizing the global energy consumption of the ventilation system. This two-level optimization architecture balances the rationality and economy of the scheduling scheme.

[0024] 6. This invention utilizes conditional generative adversarial networks (GANs) and generates samples based on geological categories, enabling targeted expansion of training data for specific geological units. Furthermore, Mahalanobis distance is used for filtering, ensuring that the synthesized samples conform to the data distribution of the geological category while avoiding the generation of anomalous or invalid data. This effectively solves the problem of insufficient training for models with small sample geological units, improves the prediction accuracy and stability of all dedicated models, and particularly enhances prediction performance under sparse geological conditions.

[0025] 7. This invention establishes a dynamic memory and combines it with importance sampling to optimize the utilization efficiency of historical samples during incremental learning. It prioritizes representative samples that are most critical to preventing model forgetting, thereby achieving optimal knowledge retention with limited storage and computing resources. This further enhances the efficiency and reliability of the online learning process, ensuring superior performance of the generated temporary prediction model.

[0026] 8. This invention employs a multi-agent reinforcement learning framework to solve the complex central optimization problem in a distributed manner. Each wind turbine agent makes decisions based on local observations and limited communication, significantly reducing computational complexity and enhancing the system's scalability and practicality in large-scale ventilation networks. The collaborative mechanism among agents enables them to spontaneously learn cooperative strategies and jointly achieve the goal of minimizing global energy consumption, demonstrating excellent swarm intelligence.

[0027] 9. This invention utilizes a cross-modal attention mechanism for feature fusion, enabling adaptive mining and weighting of the contribution of data from different sources (such as images, numerical values, and sequences) to the safety risk score. It can capture cross-modal correlation features, such as "specific rock fragment morphology accompanied by fluctuations in specific tunneling parameters," thereby generating a more comprehensive and accurate dynamic safety risk score than simple feature splicing, providing a more reliable basis for emergency decision-making.

[0028] 10. This invention embeds fluid dynamics equations as physical constraints into the reward function of reinforcement learning, essentially injecting domain knowledge into the data-driven algorithm. This guides the agent to follow basic physical laws when exploring strategies, avoiding the generation of unrealistic or fluid dynamics-violating control commands. This significantly improves the physical feasibility and engineering practicality of the learned ventilation scheduling strategy, while accelerating the algorithm's convergence speed and enhancing the stability and reliability of the entire ventilation optimization system.

[0029] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0031] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0032] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0033] In the description of this invention, the terms "lateral", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the shown orientation or positional relationship, and are only for the convenience of describing this invention and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0034] This invention provides a big data-driven method for predicting tunnel gas concentration and scheduling ventilation, which includes the following steps: S1. Based on the multi-source heterogeneous historical data of the tunnel, geological feature clustering is performed using an unsupervised learning algorithm to divide the tunnel into multiple geological prediction units; S2. For each geological prediction unit, extract multi-dimensional features containing temporal correlations, and use an ensemble learning framework to train a dedicated gas concentration prediction model to form a model library. S3. During tunnel excavation, real-time data on the operating conditions of construction machinery and rock debris images at the tunnel face are collected and integrated. A pre-trained geological identification model is used to dynamically identify the current geological category. The pre-trained geological identification model uses a ResNet-50 architecture pre-trained on the ImageNet dataset for its backbone network and is fine-tuned for rock debris image features. The model's training data comes from a large number of rock debris images collected during historical tunneling processes. These images have been labeled with specific geological category tags (such as "granite", "shale", "fault fracture zone", etc.) by geological engineers based on the lithological logging results of the corresponding sections. S4. Match the dynamically identified geological categories with the geological prediction units in the model library: If a match is found, the corresponding dedicated gas concentration prediction model is used as the effective prediction model. If the matching fails, it is determined that the area has entered an unforeseen geological anomaly zone. The emergency ventilation plan is then activated, and the online model learning process is started simultaneously. The online learning process is based on the real-time data stream newly generated in the current anomaly zone. Using the prediction model of the most adjacent geological unit as the base model, incremental learning and rapid fine-tuning are performed to generate a temporary prediction model for the anomaly zone as an effective prediction model. S5. Input real-time environmental data into an effective prediction model to obtain gas concentration prediction values ​​for future periods. Based on these prediction values, and with the goals of optimal ventilation efficiency and minimum energy consumption, generate differentiated ventilation scheduling instructions in real time through an optimization algorithm to drive the ventilation equipment in the corresponding area to execute.

[0035] The big data-driven tunnel gas concentration prediction and ventilation scheduling method involved in this invention is implemented by an integrated intelligent control system, which typically includes a data acquisition layer, a data storage and processing layer, and a decision execution layer. The data acquisition layer consists of various sensors deployed within the tunnel (such as gas sensors, image sensors, and mechanical condition sensors), responsible for collecting multi-source heterogeneous data in real time. The data storage and processing layer includes a database server and a computing server, used to store historical and real-time data, and carries all the algorithm models of this invention (such as geological clustering modules, model libraries, geological identification modules, online learning modules, ventilation optimization modules, etc.). The decision execution layer controls the operation of ventilation equipment according to the generated ventilation scheduling instructions. The steps described below are all executed collaboratively by the corresponding modules in this integrated intelligent control system.

[0036] Traditional methods for predicting tunnel gas concentrations are inaccurate due to complex and variable geological conditions. Ventilation control relies on fixed threshold responses, resulting in lag and high energy consumption. Furthermore, they lack the ability to quickly adapt and make emergency dispatch in areas with unforeseen geological anomalies.

[0037] Compared with the closest existing technologies, traditional methods are usually based on historical statistics or fixed geological models for gas prediction, which cannot dynamically adapt to geological changes. Ventilation scheduling is mainly based on post-event response, resulting in high energy consumption and low safety. In contrast, this solution achieves adaptive switching of the prediction model according to geological conditions through geological clustering, a dedicated model library, real-time geological identification and online learning. Furthermore, it generates forward-looking ventilation instructions through multi-objective optimization, fundamentally improving prediction accuracy and scheduling efficiency.

[0038] A geological prediction unit refers to a tunnel section with similar geological characteristics, defined by clustering historical geological features through unsupervised learning. The division is based on multi-dimensional features including rock strength and groundwater conditions. A dedicated gas concentration prediction model is an ensemble learning model trained for a specific geological prediction unit, capable of capturing the temporal variation of gas concentration within that unit. A model library stores prediction models corresponding to all geological prediction units for rapid matching and retrieval. A geological identification model is a pre-trained model based on deep learning, dynamically identifying the current geological category through real-time analysis of rock debris images and mechanical operating data at the tunnel face. An online learning process is a method for rapidly generating temporary prediction models in unforeseen geological anomaly areas, based on the nearest neighboring geological unit model and combined with real-time data streams for incremental learning. Ventilation scheduling instructions are ventilation equipment control strategies generated through optimization algorithms, based on predicted gas concentration values ​​and aiming for optimal ventilation efficiency and energy consumption.

[0039] This scheme first constructs a geological gas feature vector based on historical multi-source data of the tunnel, such as rock strength and harmful gas concentration. Unsupervised learning is then performed using the Toplitz inverse covariance clustering algorithm, which incorporates mandatory link constraints. These constraints are constructed from lithological logging and gas emission expert knowledge, ensuring that the clustering results are both data-cohesive and have clear engineering significance, thus dividing the tunnel into multiple geological prediction units. For each unit, multi-dimensional features with temporal correlation are extracted, and a dedicated gas concentration prediction model is trained using an ensemble learning framework and stored in a model library. During tunnel excavation, real-time images of the tunnel face's mechanical conditions and rock debris are collected. The pre-trained geological identification model dynamically identifies the current geological category and matches it with the model library. If a match is successful, the corresponding model is used to predict the gas concentration; if a match fails, it indicates the tunnel has entered an unforeseen geological anomaly zone, and an emergency ventilation plan is immediately activated. Simultaneously, using the model of the nearest neighboring geological unit as the base model, an online learning process is used to perform incremental learning by mixing replay sample sets with real-time data streams. The replay sample sets are extracted from the historical database based on data distribution similarity and given higher weights during learning to constrain the model from deviating excessively from the original knowledge, thereby quickly generating a temporary prediction model. Finally, real-time environmental data is input into an effective model to obtain predicted gas concentration values. An evaluation system is constructed that includes average wind speed, wind speed non-uniformity coefficient, peak concentration of harmful gases, and total ventilation power. Fuzzy logic is used to calculate the membership degree of each indicator, and the required air volume is determined by an improved Seagull optimization algorithm. Furthermore, the ventilation network is abstracted into a node and branch graph structure. With fan air volume and air pressure as variables, required air volume as constraint, and minimum total energy consumption as objective, a ventilation network optimization model is constructed. A deep deterministic strategy gradient reinforcement learning algorithm is used to solve the model. Through interaction between the agent and the environment, the fan coordination strategy is learned, and ventilation scheduling instructions are generated to drive equipment execution.

[0040] This solution achieves precise matching between the prediction model and geological conditions through geological clustering and a dedicated model library, significantly improving the accuracy and scenario adaptability of gas concentration prediction. The dynamic geological identification and matching mechanism ensures adaptive switching of the model during tunneling, enhancing system robustness. For unforeseen geological anomalies, the linked response of emergency ventilation and online learning provides immediate safety protection and rapid reconstruction of prediction capabilities, effectively shortening the decision-making vacuum period. Multi-objective optimization scheduling based on accurate prediction scientifically balances ventilation efficiency and energy consumption, enabling on-demand allocation of ventilation resources and reducing operating costs. The introduction of mandatory link constraints based on expert knowledge ensures that geological clustering results possess both data cohesion and engineering physical significance, enhancing the credibility of the prediction chain. The weighting strategy of historical replay samples in online learning mitigates catastrophic forgetting, ensuring the stability and generalization ability of temporary models. Multi-source information fusion and dynamic risk scoring enable proactive safety decision-making, and the two-level optimization architecture of the ventilation network balances scheduling rationality and economy. Overall, this solution promotes the intelligent and lean development of tunnel gas prediction and ventilation scheduling, effectively ensuring construction safety and improving energy efficiency.

[0041] In another technical solution, the method for predicting tunnel gas concentration and scheduling ventilation based on big data, specifically involves performing geological feature clustering through an unsupervised learning algorithm in step S1: Based on historical data such as rock strength, groundwater status, fault structure information, and historical hazardous gas concentrations, a geological-gas feature vector is constructed. The Toeplitz Inverse Covariance-based Clustering (TICC) algorithm, which introduces mandatory link constraints, is used to cluster geological-gas feature vectors. The mandatory link constraint is constructed based on expert knowledge of historical lithological logging data and gas emission patterns. It guides the algorithm to group samples with similar lithology and similar gas emission patterns into the same geological prediction unit. The specific construction process of the mandatory link constraint is as follows: experts conduct correlation analysis based on historical lithological logging data (such as rock type, mineral composition, and weathering degree) and the historical gas emission curves of the corresponding section (such as concentration fluctuation amplitude and emission rate) to identify pattern pairs with stable correspondences. For example, when the patterns of "high-strength siliceous rock" and "low-concentration stable emission" or "fragmented carbonaceous shale" and "high-concentration pulsed emission" repeatedly co-occur, they are marked as a set of mandatory link sample pairs, and a constraint set is constructed based on this to guide the clustering algorithm.

[0042] When clustering tunnel geological features using unsupervised learning, purely data-driven clustering methods may not fully reflect the inherent engineering relationship between geological genesis and gas emission behavior, resulting in insufficient consistency of the divided geological prediction units in an engineering sense, which in turn affects the accuracy of subsequent dedicated gas concentration prediction models.

[0043] Compared to the closest existing technologies, traditional geological clustering methods typically rely solely on automated data mining algorithms, lacking a mechanism to systematically integrate domain expert knowledge into the clustering process. This can lead to mathematically cohesive clustering results that are not strongly correlated with actual geological conditions and gas emission patterns. This proposed solution introduces a mandatory link constraint based on expert knowledge into the clustering algorithm, ensuring that the divided geological prediction units possess both data cohesion and clear engineering physical significance.

[0044] A geological prediction unit refers to a tunnel segment with similar geological characteristics and gas emission patterns, identified through cluster analysis. It serves as the basic unit for constructing a dedicated prediction model. A geological gas feature vector is a multidimensional numerical representation extracted from historical data to describe geological conditions and gas status. Its components include rock strength indices, groundwater inflow or level data, fault structure development parameters, and historical hazardous gas concentration monitoring values. The mandatory link constraint is a rule based on domain prior knowledge, requiring samples considered by experts to have similar lithology and gas emission patterns to be grouped into the same geological prediction unit. This constraint typically originates from the correlation analysis between historical lithology logging data and gas emission patterns. The Toplitz inverse covariance clustering algorithm is a clustering method capable of handling temporal correlations and imposing Toplitz structural constraints on the inverse covariance matrix, making it particularly suitable for processing time-dependent geological gas data.

[0045] This approach first requires collecting raw data from historical databases of tunnel construction and multi-source monitoring systems. This data includes geological survey reports, rock mechanics test records, hydrological observation logs, geological structure maps, and time-series data recorded by gas concentration sensors. Next, this multi-source heterogeneous data undergoes cleaning and standardization processes, such as normalizing rock strength, quantifying groundwater status into inflow levels, digitally encoding fault structure information, and aligning it with historical hazardous gas concentration peaks and fluctuation characteristics of the corresponding sections to construct a geological gas characteristic vector for each historical section. A crucial step lies in establishing the necessary link constraints. This requires domain experts to conduct correlation analysis between historical lithological logging data and gas concentration variation curves based on long-term engineering experience. This analysis identifies which lithological combinations typically exhibit similar gas emission behavior under specific hydrological and tectonic conditions, and marks these strongly correlated sample pairs as the necessary link sets. Subsequently, the Toplitz inverse covariance clustering algorithm was used to cluster all geological gas feature vectors. This algorithm can effectively capture common temporal dependencies in geological gas data. Furthermore, during the clustering process, the previously established mandatory link constraints were introduced as hard conditions, guiding the algorithm to perform clustering under the premise of satisfying these engineering experience constraints. Ultimately, the tunnel historical section was divided into multiple geological prediction units with high internal consistency and clear engineering significance. Samples within each unit are not only similar in geological characteristics, but their gas emission patterns are also highly correlated.

[0046] This scheme, by introducing mandatory link constraints, transforms the geological clustering process from purely mathematical calculations into a deep integration of domain experts' experience and knowledge. This ensures that the final geological prediction units not only maintain data-level cohesion but also strengthen the engineering consistency within each unit in terms of geological genesis and gas emission behavior. This lays a reliable physical foundation for training high-precision gas concentration prediction models. Compared to methods that solely rely on data-driven clustering, this scheme significantly improves the engineering rationality and interpretability of geological prediction unit partitioning, reducing prediction model bias caused by improper geological unit partitioning at its source. Because the geological conditions and gas emission patterns within each geological prediction unit are highly uniform, the dedicated gas concentration prediction model trained for each unit can more accurately capture the unique gas variation patterns of that unit, thereby improving the overall accuracy and reliability of gas concentration prediction and providing a more credible basis for subsequent ventilation scheduling decisions. The core advantage of this scheme lies in structurally integrating the experiential knowledge that is difficult for human experts to quantify into the data-driven clustering algorithm through mandatory link constraints. This achieves an effective combination of qualitative experience and quantitative analysis, enhancing the adaptability of the entire prediction system to complex geological engineering projects.

[0047] In another technical solution, the tunnel gas concentration prediction and ventilation scheduling method based on big data, in step S4, the online model learning process specifically includes: From the historical data repository, a portion of historical samples are extracted based on the similarity of data distribution with the current nearest geological unit to form a replay sample set; The replay sample set is mixed with the newly generated real-time data stream in the current anomaly region and used together as training data for incremental learning. During incremental learning and rapid fine-tuning, the replay sample set is weighted more heavily than the new sample set to constrain the model parameters from deviating excessively from the original knowledge during fine-tuning.

[0048] When learning a model online without anticipating geological anomalies, directly using real-time data streams for incremental learning may cause the model to over-adapt to new data and quickly forget most of the geological types it has learned, a problem known as "catastrophic forgetting," which affects the stability and generalization ability of the temporary prediction model.

[0049] Traditional online model update methods typically use newly acquired real-time data to fine-tune the original model. They lack organized playback and targeted protection mechanisms for historical knowledge. When the model quickly adapts to a new environment, its memory of previously learned general rules decays rapidly, leading to drastic fluctuations in prediction performance when faced with complex and ever-changing geological conditions.

[0050] The online learning process refers to a model update procedure initiated to quickly build a temporary predictive model after the system determines that it has entered an area of ​​unforeseen geological anomalies. The core of this solution lies in using a mixture of old and new data to incrementally learn a model based on the nearest neighbor geological unit. The replay sample set is a collection of historical data samples deliberately extracted from the historical data repository. These samples are not selected randomly but based on their similarity in data distribution to the current nearest neighbor geological unit. The purpose is to allow the model to review existing important knowledge relevant to the current context while learning new knowledge. Incremental learning and rapid fine-tuning are training methods in machine learning that do not require the model to learn from scratch. Instead, based on existing model parameters and combined with newly arrived data, a limited number of parameter updates are performed to enable the model to adapt to the new patterns represented by the new data without losing the memory of the original patterns. The weight allocation strategy is a method used in the incremental learning process to balance the learning intensity of old and new knowledge. By applying a higher learning weight to the replay sample set than to the new samples, the model parameters are constrained from being overly biased towards new features during updates.

[0051] In this scheme, when the geological identification model is matched with the model library and it is confirmed that the current working face is in an unforeseen geological anomaly zone, the online model learning process is immediately triggered. The system first initiates a retrieval process based on data distribution similarity from a database containing a large number of historical geological gas feature vectors. This process assesses the correlation between historical samples and the cluster centers of the current nearest geological units in the feature space, using methods such as Mahalanobis distance or cosine similarity. Then, based on the similarity scores, a certain number of the most representative historical samples are selected to form a replay sample set. Next, the system continuously collects newly generated real-time data streams from the current anomaly zone. After preprocessing and feature extraction, this data is mixed with the newly constructed replay sample set to form a balanced training dataset for subsequent model fine-tuning. In the core stage of incremental learning and rapid fine-tuning, the training algorithm applies different learning intensities to the samples in this mixed dataset. Specifically, by adjusting the weight coefficient of the error term for the replay samples in the loss function to be greater than the weight of the real-time data samples, the model incurs a greater "penalty" for prediction errors in the replay sample set with each parameter update. This forces the model to consider the general knowledge and patterns carried by historical data while striving to fit the local features presented by the new data. This fine-tuning process starts with the prediction model of the most adjacent geological unit and uses a gradient descent-based optimization algorithm for a limited number of iterations to ultimately generate a temporary prediction model for the currently unforeseen geological anomaly area.

[0052] This solution, by introducing a weighted replay sample set, effectively resists catastrophic forgetting during the online learning process. This ensures that the temporary prediction model, while rapidly adapting to local features in unknown areas, retains its predictive ability for general gas concentration changes across a wide range of geological types, thus maintaining model stability. Assigning higher learning weights to historical replay samples acts as an anchor point for model parameter updates, preventing significant drift driven by small amounts of new data. This enhances the robustness of the temporary model in the face of rapid data changes or noise, resulting in more reliable output. This mechanism enables the system to provide immediate safety by activating emergency ventilation plans when encountering geological anomalies, and also provides a predictive model with a certain degree of generalization ability within a short timeframe. This shortens the decision-making vacuum period from anomaly identification to accurate prediction recovery, improving the system's resilience in dealing with unknown risks. This solution cleverly balances the model's learning efficiency for new knowledge with the retention of old knowledge. Without significantly increasing computational complexity, it improves the overall performance of the temporary model through improved learning strategies, demonstrating the refined design of machine learning algorithms in engineering practice. Ultimately, this online learning approach enables the entire tunnel gas concentration prediction system to continuously evolve, allowing it to maintain the reliability of its core functions in unstable environments. This is crucial for ensuring long-term construction safety and ventilation efficiency.

[0053] In another technical solution, the big data-driven tunnel gas concentration prediction and ventilation scheduling method, in step S4, after determining that an unforeseen geological anomaly zone has been entered, further performs the following steps: Real-time fusion of rock debris image data from the working face, tunneling parameters of construction machinery, and information on the concentration of trace special gases collected by environmental sensors; Based on the fused information, a pre-trained deep learning model is used to calculate the dynamic safety risk score of the current working surface; When the safety risk score exceeds a preset threshold, a high risk is identified, and the emergency ventilation plan's level and response measures are adjusted accordingly. The preset threshold is set at multiple levels based on historical safety accident data and expert experience. For example: a dynamic safety risk score below 60 is considered "low risk," requiring only enhanced monitoring; a score between 60 and 80 is considered "medium risk," requiring localized enhanced ventilation; and a score above 80 is considered "high risk," requiring immediate full-section strong wind evacuation and suspension of tunneling operations. The emergency ventilation plan's levels correspond one-to-one with these risk ranges.

[0054] In areas with unforeseen geological anomalies, traditional single-gas concentration monitoring methods cannot comprehensively and proactively perceive integrated safety risks, leading to delayed emergency response and an inability to achieve precise and dynamic adjustments to ventilation plans.

[0055] Traditional methods typically rely on a single gas concentration threshold for passive response, lacking the fusion analysis of multi-source information from the working face. This makes it impossible to provide early warnings when other risk factors (such as abnormal rock fragment morphology or sudden changes in tunneling parameters) occur even when gas concentrations are within acceptable limits. This proposed solution integrates rock fragment images, mechanical operating conditions, and trace gas information, and utilizes a deep learning model to calculate a dynamic safety risk score. This enables comprehensive and proactive perception of the safety status of unforeseen geological anomalies and precise dynamic adjustment of ventilation plans.

[0056] Dynamic safety risk scoring is a quantitative assessment value calculated based on real-time fusion of multi-source information using a pre-trained deep learning model. It comprehensively reflects the overall safety status of the current working face. Rock debris image data refers to visual information collected by camera equipment installed at the working face, reflecting the shape, size, edges, and surface features of broken rock fragments. Tunneling parameters of construction machinery refer to real-time data generated by tunneling equipment during operation, including but not limited to mechanical operating data such as tunneling speed, cutterhead torque, thrust, and vibration frequency. Trace gas concentration information refers to data on harmful gases detected by high-precision gas sensors, where the concentration does not reach the conventional alarm threshold but shows abnormal fluctuations or specific combinations. Emergency ventilation plan levels and response measures refer to pre-established ventilation control strategies for different risk levels, including the number of fans started, air volume, wind direction control, and associated evacuation or work restriction measures.

[0057] In this technical solution, once the system determines that it has entered an area of ​​unforeseen geological anomalies, it immediately initiates a multi-source information fusion and dynamic risk assessment process. First, the system collects real-time data in three parallel aspects: continuously acquiring images of rock debris using high-definition explosion-proof cameras deployed at the tunnel face; collecting tunneling parameters using sensors installed on the tunneling machinery; and monitoring changes in the concentration of trace special gases in the environment using a high-precision gas sensor array. All of this data is synchronously transmitted to the central processing unit. Next, a multimodal fusion deep learning model is constructed to process this heterogeneous data. This model includes multiple feature extraction branches: for rock debris image data, a convolutional neural network is used to automatically learn and extract visual features related to geological stability, such as the grain size distribution, edge sharpness, and surface texture of the rock debris; for numerical data such as tunneling parameters, a fully connected layer is used to learn its deep features and capture abnormal patterns in the mechanical stress state; for time-series data such as trace special gas concentration information, a time-series encoder is used to capture the dynamic trends and abnormal fluctuations in its concentration changes. Then, a cross-modal attention mechanism is used to deeply fuse these features from different sources. This mechanism can adaptively weigh the contribution of different data sources to the current risk state. For example, it can identify the extremely high risk implied by the combination of features such as "specifically sharp rock cuttings," "sudden drop in tunneling torque," and "the presence of trace amounts of acetylene gas." The fused features are finally input into a regression layer, outputting a comprehensive dynamic safety risk score. The system compares the calculated dynamic safety risk score with multiple preset risk thresholds. When the score exceeds the minimum threshold, the system issues a primary alarm; when the score enters the high-risk range, the system not only triggers audible and visual alarms but also automatically adjusts the level of the emergency ventilation plan, such as upgrading from "enhanced monitoring" to "local enhanced ventilation," or further upgrading to "full-section strong wind evacuation," and adjusting corresponding response measures accordingly, such as advising workers to evacuate or adjusting the tunneling pace. The entire process forms a closed-loop active safety control system from multi-source perception and risk calculation to plan adjustment.

[0058] This solution integrates multi-dimensional information such as rock debris morphology, mechanical conditions, and gas concentration. The system can capture early risk signs that a single gas sensor cannot identify, enabling proactive perception of potential hazards and shifting safety management from reactive response to proactive early warning. Based on dynamic risk scoring, the system automatically adjusts ventilation plans, avoiding the lag of acting only when gas concentration exceeds limits and preventing energy waste caused by blindly increasing airflow, thus achieving precise allocation of emergency resources. The application of a cross-modal attention mechanism allows the system to intelligently capture the inherent correlations between different information sources, such as identifying the coupling relationship between specific lithological conditions, mechanical load, and gas escape patterns, thereby generating more comprehensive and accurate risk assessment results than simple rule-based judgments. This dynamic scoring mechanism greatly enhances the safety management capabilities of unforeseen geological anomalies, providing reliable decision support for safe tunneling under complex geological conditions and significantly reducing the probability of safety accidents in high-risk geological sections.

[0059] In another technical solution, the big data-driven tunnel gas concentration prediction and ventilation scheduling method, in step S5, takes optimal ventilation efficiency and minimum energy consumption as multiple objectives, and generates differentiated ventilation scheduling instructions in real time through an optimization algorithm, which is specifically achieved through the following steps: S5.1 Construct a comprehensive evaluation index system that includes the average wind speed in the tunnel, the wind speed non-uniformity coefficient, the peak concentration of harmful gases, and the total power of the ventilation system; calculate the membership degree value of each individual index based on the fuzzy logic method; use the improved Seagull optimization algorithm to optimize the membership degree value of each index to obtain the equilibrium solution that makes the membership degree value of all indices optimal as a whole, and determine the required air volume of each branch in the ventilation network to meet the predicted gas concentration value based on the equilibrium solution. S5.2. The tunnel ventilation network is abstracted as a graph structure composed of nodes and branches, where nodes represent tunnel faces, junctions, or ventilation openings, and branches represent tunnel sections. The wind resistance characteristics of each branch are recorded. The wind resistance characteristics of each branch in the ventilation network are obtained through a combination of field measurements and fluid dynamics simulation. Specifically, in the early stage of tunnel construction, the initial wind resistance coefficient is obtained by adjusting the fans and measuring the wind speed and wind pressure of each branch, and then using the Darcy-Weisbach formula to back-calculate the initial wind resistance coefficient. As tunneling progresses, the wind resistance coefficient is adjusted based on the tunnel support type, cross-sectional dimensions, and rock wall roughness. Roughness is dynamically updated and corrected using computational fluid dynamics simulation to optimize the drag coefficient. A ventilation network optimization model is constructed using the airflow and pressure of all fans in the network as optimization variables, the required airflow obtained in step S5.1 as a constraint, and minimizing the total energy consumption of the fans as the optimization objective. The ventilation network optimization model is solved using a deep deterministic policy gradient reinforcement learning algorithm. Through continuous interaction between the agent and the ventilation network environment, the fan collaborative control strategy that minimizes total energy consumption is learned, and ventilation scheduling instructions are generated.

[0060] Traditional tunnel ventilation scheduling methods struggle to simultaneously optimize ventilation efficiency and reduce system energy consumption. They lack a scientific multi-index collaborative decision-making mechanism and cannot quickly solve for the globally optimal fan control strategy in complex ventilation network environments.

[0061] Traditional ventilation scheduling often relies on manual experience or simple single-objective control, such as adjusting fans based solely on gas concentration thresholds at a specific location. This approach struggles to comprehensively consider multiple conflicting indicators, including wind speed uniformity within the tunnel, peak concentrations of harmful gases, and total system power. Furthermore, it fails to efficiently handle the complex coupling relationships between fans in the ventilation network, leading to poor ventilation or significant energy waste. This solution, by constructing a comprehensive evaluation index system and combining it with improved optimization algorithms and reinforcement learning, achieves two-level decision-making: scientific allocation of ventilation demand and coordinated optimization of fan energy consumption.

[0062] The comprehensive evaluation index system is a set of quantitative indicators used to comprehensively measure the operating status of a ventilation system. It typically includes average wind speed reflecting the overall airflow, wind speed non-uniformity coefficient assessing airflow stability, peak concentrations of harmful gases related to safety, and total power of the ventilation system reflecting economic efficiency. Fuzzy logic is a mathematical tool for handling uncertainty and subjective judgment. It can transform the actual measured values ​​of these indicators into a unified membership value representing the degree of superiority or inferiority, thus solving the problem of direct comparison between indicators of different dimensions. The improved seagull optimization algorithm is a swarm intelligence optimization algorithm that simulates the migration and attack behavior of seagulls. In this invention, it has been adaptively improved to find the solution in a complex membership space that achieves the optimal balance of all ventilation efficiency indicators, and based on this, determines the required air volume for each branch in the ventilation network. The ventilation network optimization model abstracts the actual tunnel ventilation system into a graph structure composed of nodes and branches. Nodes represent key locations such as tunnel faces, junctions, or ventilation openings, while branches represent tunnel sections connecting these nodes, recording their wind resistance characteristics. This model uses the airflow and pressure of all fans as adjustable decision variables, with the required airflow to meet safety requirements as a hard constraint, and aims to minimize the total system energy consumption. Deep deterministic policy gradient reinforcement learning (DPRL) is an advanced artificial intelligence method that learns optimal strategies from interactions with the environment. In this scheme, it is used to solve the ventilation network optimization model. An agent (which can be understood as a fan controller) continuously tries different fan combinations and observes the resulting energy consumption and airflow distribution effects, ultimately autonomously learning the fan cooperative control strategy that minimizes total energy consumption while meeting airflow requirements.

[0063] This solution first enters the first stage of ventilation scheduling instruction generation, namely the optimized allocation of required air volume. The system acquires real-time data on the average wind speed, wind speed non-uniformity coefficient, peak concentration of harmful gases, and the total power of the ventilation system within the tunnel. Next, fuzzy logic is used to calculate the membership degree value for each indicator, which ranges between 0 and 1, with values ​​closer to 1 indicating better performance. Then, an improved seagull optimization algorithm is employed to handle this multi-objective optimization problem. This algorithm simulates the intelligent behaviors of seagull populations, such as flight, gathering, and attack, and performs efficient parallel search in the membership degree space. Its improvement lies in the introduction of a dynamic adaptive adjustment mechanism to balance global exploration with local development capabilities, avoiding premature entrapment in local optima. It outputs an equilibrium solution that ensures all ventilation efficiency indicators reach an overall satisfactory level. Based on this solution, the system can calculate the required air volume for each branch in the ventilation network to meet the predicted gas concentration values ​​for future periods. The second stage then begins: the collaborative optimization of fan energy consumption. The system abstracts the tunnel ventilation network into a graph structure, clearly defining all nodes and branches, and recording or updating the wind resistance characteristics of each branch in real time via sensors. Based on this, a ventilation network optimization model is constructed, using the required air volume obtained in the first stage as a constraint that must be met, and minimizing the total energy consumption of all fans in the network as the optimization objective. To solve this complex, nonlinear optimization model, the system employs a deep deterministic policy gradient reinforcement learning algorithm. In this process, each fan can be considered an agent, interacting with the environment (i.e., the ventilation network). The agent outputs the fan's air volume and pressure adjustment actions, then observes the energy consumption and air volume satisfaction after executing the action, and evaluates the merits of the action according to a carefully designed reward function (e.g., meeting the air volume constraint yields a positive reward, while increasing energy consumption yields a negative reward). Through extensive trial and error and learning, the agent ultimately masters a collaborative control strategy—a set of intelligent solutions for coordinating the operating parameters of each fan to minimize the total energy consumption of the system while ensuring the required air volume. Finally, the system decodes this strategy into specific, executable ventilation scheduling instructions and sends them to the corresponding ventilation equipment.

[0064] This solution constructs a comprehensive evaluation index system and uses fuzzy logic and improved optimization algorithms for solution. The system can scientifically balance multiple potentially contradictory indicators within ventilation efficiency, such as ensuring that harmful gas concentrations do not exceed limits while also considering wind speed uniformity. This allows for the determination of a more reasonable and accurate demand air volume, laying a reliable foundation for subsequent energy consumption optimization. Leveraging the advantages of deep deterministic policy gradient reinforcement learning algorithms in handling complex nonlinear optimization problems and sequential decision-making, the system can directly optimize fan operation strategies, fully tapping the energy-saving potential of the ventilation network. This avoids energy waste caused by over-ventilation or mutual cancellation between fans, which may occur in traditional methods, achieving global energy consumption minimization under safety constraints. This two-level optimization architecture, which decouples demand-side air volume allocation from supply-side fan coordination, ensures both the safety and rationality of ventilation effects while improving the economic efficiency of system operation, effectively coordinating safety and energy efficiency goals. This solution enhances adaptability to complex ventilation network environments. By autonomously learning the optimal control strategy through data-driven methods, it reduces reliance on precise physical models, improving the intelligence level and reliability of the ventilation scheduling system under varying conditions.

[0065] In another technical solution, the tunnel gas concentration prediction and ventilation scheduling method based on big data, in step S1, the fusion process of multi-source heterogeneous historical data further includes a data augmentation step, which specifically involves: A conditional generative adversarial network is used to generate synthetic samples by using geological category labels obtained by clustering through the Topplitz Inverse Covariance Clustering (TICC) algorithm as conditional input. Calculate the Mahalanobis distance between the synthetic sample and the cluster center of its geological category, compare the Mahalanobis distance with a preset threshold, and retain the synthetic samples whose Mahalanobis distance is not greater than the preset threshold to obtain the filtered synthetic samples. The selected synthetic samples were added to the geological-gas feature vector set to expand the training dataset used to train the dedicated gas concentration prediction model.

[0066] When training a dedicated gas concentration prediction model based on geological clustering results, the historical data sample size of some geological prediction units is too small, resulting in insufficient model training, low prediction accuracy and poor stability. This problem is particularly prominent under geological conditions with sparse data.

[0067] Traditional data augmentation methods typically employ simple oversampling or the addition of random noise. The resulting synthetic samples often lack specificity and fail to accurately reflect the true data distribution of specific geological categories. Sometimes, they may even introduce invalid or misleading samples, thus failing to effectively improve the training quality of small-sample geological unit models. This proposed solution combines conditional generative adversarial networks (GANs) and a Mahalanobis distance-based selection mechanism to achieve targeted and high-quality augmentation of training data for specific geological units.

[0068] Conditional Generative Adversarial Networks (GANs) are deep learning models consisting of a generator and a discriminator. Their key feature is the ability to generate data that meets specific requirements based on given conditional information. In this invention, this conditional information is the geological category label. The geological category label is a category identifier assigned to each geological prediction unit after clustering historical data using the Toplitz inverse covariance clustering algorithm. Synthetic samples refer to new data samples generated by the generator part of the GAN, simulating the distribution of real data. Mahalanobis distance is a metric for measuring the distance between a sample point and a data distribution. It considers the correlation between different dimensions of the data and therefore reflects the true location of the sample in the distribution better than Euclidean distance. The cluster center is a benchmark point in the feature space representing the average feature of all samples within a geological prediction unit. The preset threshold is a pre-set distance value used to determine whether the synthetic sample is sufficiently close to the core distribution area of ​​its geological category.

[0069] After clustering the geological features of historical multi-source heterogeneous data to form multiple geological prediction units, the system initiates a data augmentation process to expand the initial dataset used to train the dedicated gas concentration prediction model. First, the system uses the geological category label and its corresponding geological gas feature vector for each geological prediction unit as training data, inputting them into a conditional generative adversarial network (GAN). During training, the generator learns to receive a random noise vector and a specific geological category label, attempting to generate synthetic samples that match the features of that geological category; the discriminator learns to distinguish whether the input sample is real historical data or a synthetic sample generated by the generator, while also receiving the geological category label as a judgment condition. Through repeated iterations of adversarial training, the generator eventually produces synthetic samples that are highly similar to the distribution of real data and possess the characteristics of the specified geological category. Next, the system performs quality screening on all generated synthetic samples. For each synthetic sample, the Mahalanobis distance between it and the cluster center of its own geological category is calculated. This distance quantifies the degree to which the synthetic sample deviates from its category core distribution in the feature space. Then, the calculated Mahalanobis distance is compared with a preset threshold. Only synthetic samples whose Mahalanobis distance is no greater than a preset threshold—that is, those falling within or very close to the typical characteristic regions of this type of geological unit—are retained; samples that are too far away, potentially representing anomalies or invalid patterns, are discarded. This screening step is crucial, ensuring the quality and representativeness of the augmented data. Finally, the system adds the high-quality synthetic samples retained after screening to the original geological gas feature vector set, forming a larger and more diverse training dataset. This augmented dataset is then used to train the dedicated gas concentration prediction model for each geological prediction unit mentioned in step S2.

[0070] This approach generates samples based on geological categories, providing targeted data augmentation that effectively compensates for data scarcity in specific geological units, especially those with sparse samples. This ensures that the training of each dedicated model is built upon a more robust data foundation. Mahalanobis distance is used for selection, guaranteeing that the generated synthetic samples not only increase in quantity but also match the core data characteristics of their corresponding geological categories. This avoids introducing meaningless or noisy samples that could interfere with model training, thereby improving the model's prediction accuracy and stability. This method significantly improves the performance of prediction models built under sparse geological conditions, enabling relatively reliable predictions even for rare or poorly explored geological types. It enhances the adaptability and robustness of the entire gas concentration prediction system under different geological environments. Ultimately, by improving the training quality of all dedicated models, it provides strong support for the reliability of the entire tunnel gas concentration prediction and ventilation scheduling chain from a fundamental data perspective, laying a more solid data foundation for achieving precise control of safety and energy efficiency.

[0071] In another technical solution, the big data-driven tunnel gas concentration prediction and ventilation scheduling method, in the online model learning process of step S4, the specific process of incremental learning and rapid fine-tuning includes: establishing a dynamically updated memory to store representative historical samples; when performing incremental learning, selecting samples from the memory through an importance sampling strategy and mixing them with the real-time data stream.

[0072] During the online learning process of the model, how to efficiently utilize limited computing and storage resources to select and reuse historical samples in order to alleviate the "catastrophic forgetting" of old knowledge when the model adapts to new data, and ensure that the temporary prediction model can maintain its predictive ability for existing geological types while quickly learning new features.

[0073] Traditional incremental learning methods typically mix recent real-time data with randomly selected or all historical data when encountering unforeseen geological anomalies. This approach fails to distinguish the importance of historical samples, potentially leading to decreased generalization ability due to the loss of key historical knowledge or inefficient learning due to interference from invalid samples. This proposed solution optimizes the utilization efficiency of historical samples by introducing a dynamically updated memory and combining it with an importance sampling strategy, aiming to achieve the best knowledge retention effect with limited resources.

[0074] The dynamically updated memory is a storage structure specifically designed for online learning processes. It doesn't simply cache all historical data, but selectively stores historical samples deemed representative during the feature space or learning process. These samples typically reflect typical gas concentration variation patterns or key geological features of their respective geological prediction units. The importance sampling strategy is a probabilistic sample selection method that assigns a probability of selection to each historical sample based on its importance to the current learning task. Samples with higher importance are more likely to be selected for model fine-tuning, thus guiding the model to prioritize consolidating these key knowledge points during updates. The determination of representative historical samples relies on predefined criteria. For example, a sample might be representative because its feature vector is close to the cluster center of the geological prediction unit, or it might carry special information because it was previously difficult for the model to predict accurately during historical training.

[0075] In this scheme, when the system initiates the online learning process to address unforeseen geological anomalies, the construction and use of the dynamic memory database unfolds simultaneously. First, the system selects an initial batch of samples from a database containing a large number of historical geological gas feature vectors, based on preset representativeness criteria, and loads them into the memory database. These criteria may include: the distance between a sample and the cluster center of its geological unit, the magnitude of the loss function value of the sample during historical training, or the rarity of certain features of the sample. During subsequent tunnel excavation, as real-time data from the current anomaly area continues to flow in, the system maintains the memory database in parallel. Maintenance operations include: periodically evaluating the representativeness of existing samples in the memory database; when newer, more representative samples appear, weaker representative samples may be replaced; or, based on the model's mastery of various historical knowledge during recent incremental learning, dynamically adjusting the retention ratio of various types of samples in the memory database. When incremental learning and rapid fine-tuning are required, the system does not throw all samples from the memory database into training at once, but instead initiates an importance sampling strategy. This strategy calculates an importance score for each sample in the memory, based on factors such as the sample's similarity to the current real-time data distribution, the sample's own uncertainty, or the frequency of its historical usage. Subsequently, the system extracts a certain number of samples from the memory based on the probability distribution formed by these importance scores, creating a replay sample set for the current incremental learning. This replay sample set is then mixed with the currently acquired real-time data stream to form the training dataset. During model fine-tuning, the weights of the loss function for these historical samples selected through importance sampling are typically set higher than for ordinary new samples, emphasizing the consolidation of key historical knowledge. Through this refined sample management and selection mechanism, the system maximizes the retention of the essential components of historical data within limited storage space and computation cycles, thereby generating a higher-performing temporary prediction model.

[0076] This scheme combines a dynamic memory bank with importance sampling. The system intelligently identifies and prioritizes the retention of historical samples most critical to preventing model forgetting. This allows the online learning process to more effectively mitigate catastrophic forgetting under the same resource conditions, ensuring that the temporary prediction model can quickly adapt to new environments while maintaining its predictive ability for learned patterns, thus improving model robustness. The importance sampling mechanism ensures that limited computational resources are concentrated on learning the most valuable historical information, avoiding redundant computation on a large number of ordinary samples. This not only improves the efficiency of the incremental learning process and shortens the generation time of the temporary model, but also indirectly reduces the system's computational load. Because the training data includes carefully selected representative samples covering different geological models, the temporary prediction model exhibits better stability and generalization ability when outputting gas concentration predictions, enhancing the reliability of its prediction results for ventilation scheduling decisions. This design endows the entire gas concentration prediction system with a continuous and robust evolutionary capability, enabling it to continuously accumulate and integrate knowledge through efficient online learning when encountering unknown geological environments, thereby maintaining high-precision prediction performance in the long term. This has profound significance for ensuring the safety and energy efficiency control of tunnel construction.

[0077] In another technical solution, the method for predicting tunnel gas concentration and scheduling ventilation based on big data, in step S5.2, specifically includes solving the ventilation network optimization model by modeling the fan cooperative control problem as a partially observable Markov decision process, where each fan is an agent, and a multi-agent reinforcement learning algorithm is used for distributed training, and local observation information is shared through the communication mechanism between agents.

[0078] When implementing multi-fan coordinated control in complex ventilation networks, traditional centralized optimization methods face problems such as excessive computational complexity and poor scalability. Furthermore, a single controller cannot fully perceive the real-time status of each node in the network, making it difficult to guarantee the real-time performance and global optimality of the control strategy.

[0079] Traditional ventilation network control typically employs a central controller for unified calculation and decision-making based on global information. As the ventilation network expands and the number of fans increases, the computational burden of this centralized optimization increases dramatically, resulting in significant decision delays and making it difficult to meet real-time scheduling requirements. Furthermore, a failure of the central controller could paralyze the entire system. This solution models the fan cooperative control problem as a partially observable Markov decision process and uses multi-agent reinforcement learning for distributed training, achieving decentralized cooperative decision-making and effectively improving the system's scalability and robustness.

[0080] In this scheme, the partially observable Markov decision process (MAD) is a mathematical framework used to describe distributed decision-making problems. Within this framework, each agent, i.e., each wind turbine, cannot fully grasp all the state information of the entire ventilation network, but can only observe local environmental information relevant to itself, such as wind speed, wind pressure, and gas concentration at neighboring nodes in its branch. This characteristic of local observation better reflects the actual situation of sensor deployment in practical engineering. The multi-agent reinforcement learning algorithm is a machine learning method suitable for multi-decision-maker environments. In this method, each wind turbine acts as an independent agent, interacting with the ventilation network environment and autonomously learning how to adjust its airflow and wind pressure output based on its local observations. The communication mechanism between agents refers to the channels and protocols that allow these distributed wind turbine agents to exchange limited information, such as exchanging their observed local wind speed or wind pressure data, thereby compensating for the limitations of a single agent's perspective and providing an information basis for collaborative decision-making.

[0081] In this scheme, during the solution phase of the ventilation network optimization model, the system first abstracts the entire ventilation network into a graph structure composed of nodes and branches, and defines each fan in the network as an independent agent. Each fan agent has the ability to perceive its local environment, continuously acquiring wind speed and wind pressure sensor readings near its installation location, as well as communication information that may be received from neighboring agents. The cooperative control problem of the entire fan group is formally modeled as a partially observable Markov decision process. At each decision moment, each agent independently makes a control action based on its own local observations, i.e., adjusting its airflow or wind pressure settings. The actions of all these agents collectively affect the ventilation network environment, changing the environmental state and feeding back a new local observation and a shared global reward signal, such as the system's total energy consumption. A distributed training method is used to train all these fan agents. During the training process, the agents continuously optimize their decision strategies through trial and error. The key lies in the fact that agents share their local observation information through a pre-defined communication mechanism. For example, a wind turbine agent located at a network branch can broadcast its perceived risk of airflow blockage to upstream wind turbine agents. This limited information sharing allows each agent to consider not only its own state but also, to some extent, infer the global state of the network when making decisions. Through numerous rounds of iterative training, the wind turbine agent swarm eventually learns a highly coordinated control strategy. That is, each agent knows how to adjust its operating parameters based on local observations and received neighbor information in order to cooperate with other agents and jointly achieve the global goal of minimizing the system's total energy consumption.

[0082] This scheme decomposes the complex central optimization problem into distributed decision-making by multiple agents, effectively distributing the computational burden across individual fan controllers. This avoids the computational bottleneck of a single control center, enabling the system to easily handle the optimization control of large-scale ventilation networks and significantly improving the scalability and practicality of the method. Each fan agent possesses independent decision-making capabilities. Even if some agents fail or their communication links are interrupted, the remaining agents can still operate based on their existing strategies and remaining local information, ensuring the maintenance of the basic functions of the entire ventilation system under abnormal conditions and enhancing the system's robustness. Agents share local information through communication, enabling the entire system to exhibit efficient global collaborative behavior in a bottom-up manner. Without central commands, fans can self-organize and adjust their operating states to jointly achieve energy-saving goals, demonstrating strong swarm intelligence. This decision-making mode based on local perception and limited communication more closely resembles the sensor deployment conditions of actual engineering projects, reducing the system's dependence on a globally accurate model. The learned control strategies naturally possess stronger engineering feasibility and can adapt more quickly to changes in ventilation network topology or wind resistance characteristics.

[0083] In another technical solution, the calculation process of dynamic safety risk score in the tunnel gas concentration prediction and ventilation scheduling method based on big data further includes: constructing a multimodal fusion module, extracting features from rock debris image data, tunneling parameters and trace special gas concentration information through convolutional neural networks, fully connected layers and temporal encoders respectively, and using a cross-modal attention mechanism to fuse these features.

[0084] When calculating dynamic safety risk scores, how can we effectively integrate different types of data such as rock debris images, tunneling parameters, and trace gas concentrations to avoid information loss or neglect of correlation that may be caused by simple feature splicing in traditional methods, thereby improving the accuracy and comprehensiveness of risk perception?

[0085] Traditional methods for processing multi-source safety monitoring data typically employ simple feature stitching or weighted averaging. This approach struggles to capture the deep, non-linear relationships between different modalities (such as visual features of images, numerical sequences of mechanical parameters, and temporal variations in gas concentrations). For instance, it may overlook the extremely high risk implied by the simultaneous occurrence of "specific sharp rock cutting morphology," "a sudden surge in tunneling torque," and "anomaly in trace amounts of acetylene gas." This proposed solution constructs a multi-modal fusion module and introduces a cross-modal attention mechanism to adaptively learn and weigh the contributions of different data source features to the overall risk score, thereby achieving a more accurate and reliable safety status assessment.

[0086] The multimodal fusion module is a specialized deep learning architecture responsible for receiving and processing data from different sources and types. In this scheme, its inputs mainly include rock debris image data, tunneling parameters of construction machinery, and information on the concentration of trace special gases in the environment. Convolutional neural networks (CNNs) are a type of deep learning model particularly well-suited for processing image data. Through their convolutional layers, pooling layers, and other structures, they can automatically extract spatial features related to geological stability and risk from rock debris images, such as the grain size distribution, edge sharpness, and surface texture of the rock debris. Fully connected layers are a basic layer structure in neural networks, where each neuron is connected to all neurons in the previous layer. They are commonly used to process structured numerical data, such as converting numerical inputs like tunneling parameters into deeper feature representations. Temporal encoders are a type of neural network model used to process time-series data, such as recurrent neural networks or their variants. They can capture the dependencies and dynamic trends of data over time, and are particularly suitable for analyzing fluctuation patterns in the concentration of trace special gases. The cross-modal attention mechanism is the core component of this invention to achieve effective fusion. It is a computational mechanism that mimics human attention allocation and can dynamically calculate the correlation weights between features of different modalities, thereby determining which feature combinations should be given more attention during the fusion process.

[0087] In this scheme, the calculation of dynamic safety risk score begins with the synchronous acquisition and feature extraction of multi-source heterogeneous data. The system continuously acquires images of rock debris through explosion-proof camera equipment deployed at the tunnel face. This raw image data is fed into a pre-trained convolutional neural network. This network gradually extracts visual features from low to high levels through multi-layer convolution and pooling operations. These features constitute a feature vector that can characterize the current rock debris morphology. At the same time, the system reads real-time tunneling parameters from the tunneling equipment sensors, such as tunneling speed, cutterhead torque, and thrust. These numerical sequences are first preprocessed, such as normalization, and then input into a sub-network composed of multiple fully connected layers. The fully connected layers use nonlinear transformations to map these raw mechanical operating parameters to a feature space that better reflects their deeper patterns, capturing abnormal patterns such as "torque drop". For trace amounts of special gas concentration information acquired through a high-precision gas sensor array, this typical time-series data is processed by a time-series encoder. Temporal encoders, such as Long Short-Term Memory (LSTM) networks, read gas concentration sequences step-by-step and utilize their internal memory mechanisms to capture dynamic trends and potential periodic or sudden fluctuations in concentration changes. After extracting feature vectors from image, numerical, and time-series data respectively, the system combines these feature vectors into a preliminary multimodal feature set and feeds it into a cross-modal attention mechanism for deep fusion. This attention mechanism calculates a correlation score for each feature pair (e.g., a rock texture feature and a torque fluctuation feature), representing the importance of the association between these two features from different sources in assessing the current overall risk. Based on these scores, the attention mechanism recalibrates and weights the contributions of each feature vector, thereby generating a unified contextual feature representation that incorporates cross-modal information. This feature representation, rich in cross-modal correlation information, is finally input into a prediction layer, typically one or more fully connected layers, outputting a comprehensive, quantified dynamic safety risk score.

[0088] This solution overcomes the limitations of traditional simple splicing or weighted averaging methods, automatically discovering and utilizing the complex, non-linear intrinsic relationships between features from different sources. This allows for a more comprehensive characterization of the working face's safety status. For example, it can identify potential risks implied by the simultaneous appearance of multiple weak anomalies, which are often imperceptible from a single data source. By dynamically calculating the importance weights between features, the system adaptively focuses on the feature combinations most relevant to the current risk state when fusing information, suppressing interference from irrelevant or redundant information. This results in a more accurate safety risk score and greater sensitivity to early signs of risk, achieving true proactive awareness. This in-depth feature fusion significantly enhances the overall assessment capability of the safety status of unforeseen geological anomalies, providing a more reliable and solid basis for timely adjustment of emergency ventilation plans and the implementation of targeted response measures under complex geological conditions, thereby significantly improving the safety management level of high-risk geological sections.

[0089] In another technical solution, the big data-driven tunnel gas concentration prediction and ventilation scheduling method further introduces physical constraints in step S5.2 when constructing the ventilation network optimization model. The Darcy-Weisbach equation and continuity equation in fluid mechanics are embedded as constraints into the reward function of the deep deterministic policy gradient reinforcement learning algorithm.

[0090] When using a deep deterministic gradient reinforcement learning algorithm to solve the ventilation network optimization model, the purely data-driven learning process may produce fan control strategies that violate the basic principles of fluid mechanics. This can lead to the generation of ventilation scheduling commands being unable to be effectively executed in the actual ventilation network, affecting the stability and reliability of the system.

[0091] Traditional ventilation optimization control methods either rely on simplified linear models that ignore the inherent nonlinear characteristics of fluid networks, or depend entirely on data-driven reinforcement learning algorithms. However, the agent lacks guidance from physical laws when exploring strategies, potentially resulting in physically infeasible fan control actions that could even disrupt the stability of the ventilation network. This solution embeds the Darcy-Weisbach equation and the continuity equation as physical constraints into the reward function of the reinforcement learning algorithm, injecting domain knowledge into the data-driven approach and ensuring the physical rationality of the learning strategy.

[0092] In this scheme, the ventilation network optimization model is a mathematical model that abstracts the tunnel ventilation system into a graph structure. It uses the airflow and pressure of all fans in the network as optimization variables, seeking to minimize the total energy consumption of the fans while satisfying the required airflow. The deep deterministic policy gradient reinforcement learning algorithm is a reinforcement learning method applicable to continuous action spaces. It learns the optimal policy through the interaction between the agent and the environment. Its core components include a policy network for decision-making, a value network for evaluating the state-action value, and a reward function to guide the agent's learning. The Darcy-Weisbach equation is a core equation in fluid mechanics describing the pressure drop caused by frictional resistance when fluid flows in a pipe. It establishes a quantitative relationship between the wind resistance characteristics of the airflow path, the airflow volume, and the resulting pressure drop. The continuity equation is the embodiment of the law of conservation of mass in fluid mechanics. When applied to the ventilation network, it requires that the total airflow into any node must be equal to the total airflow out of that node, thus ensuring the rational distribution of airflow in the network. Physical constraint embedding refers to incorporating the aforementioned basic equations of fluid mechanics as necessary conditions into the design of the reward function of the reinforcement learning algorithm, so that the behavior of the agent is implicitly constrained by physical laws.

[0093] When constructing and preparing to solve the ventilation network optimization model, the system first abstracts the ventilation network into a graph structure composed of nodes and branches based on the physical layout of the tunnel. The connection relationships between each node and branch are clearly defined, and the drag coefficient of each branch is recorded or acquired in real time via sensors. Subsequently, the system initiates the training process based on a deep deterministic policy gradient reinforcement learning algorithm. A key innovative step in this process is the careful design of the reward function to incorporate physical constraints. In addition to a negative reward term reflecting the total energy consumption of the system, the reward function also includes two penalty terms based on physical equations. The first penalty term originates from the Darcy-Weisbach equation. The system calculates in real time the pressure drop theoretically expected by the equation for each branch in the ventilation network under its current airflow, based on its drag coefficient. This theoretical pressure drop is then compared with the actual wind pressure setpoint output by the fan strategy network acting on that branch. The difference between the two, i.e., the degree of mismatch between the wind pressure setpoint and the value calculated by physical laws, is calculated and used as a penalty term; the higher the degree of mismatch, the greater the penalty. The second penalty term originates from the continuity equation. The system examines each node in the network, calculating the algebraic sum of the airflow entering and leaving that node, which theoretically should be zero. The actual calculated node airflow imbalance is used as another penalty term; the more unconserved the node airflow, the greater the penalty. These two physical penalty terms, together with the reward term representing energy consumption, constitute a new composite reward function. During the interactive training between the reinforcement learning agent and the ventilation network environment, the agent's policy network outputs the fan's airflow and pressure actions, and the environment provides the agent with a new state and a reward value calculated based on this composite reward function. The agent's goal is not only to maximize the cumulative reward (i.e., minimize total energy consumption) but also to minimize the penalty for violating the physical equations. Through thousands of such trial-and-error iterations, the agent's policy network gradually learns how to adjust the fan's operating parameters while satisfying the basic laws of fluid mechanics, thereby minimizing the system's total energy consumption while ensuring airflow requirements. This training method is equivalent to injecting a physical guide into the data-driven learning process, ensuring that the exploration process is always constrained within a physically feasible solution space.

[0094] This approach significantly improves the engineering practicality and reliability of the learned ventilation scheduling strategy. Because the strategy is guided and filtered by physical laws during training, the generated fan control commands closely match the actual fluid dynamics characteristics of the ventilation network, avoiding impractical or ineffective adjustments and ensuring ventilation effectiveness. This domain-knowledge-integrated method reduces reliance on extensive trial and error, allowing the agent to explore physically feasible directions, thus accelerating the convergence speed of the reinforcement learning algorithm and improving training efficiency. Introducing physical constraints also enhances the stability and robustness of the entire ventilation optimization system in the face of uncertainty and disturbances, as its core decision-making mechanism is based on universal physical laws, rather than purely relying on potentially incomplete or noisy data. Fundamentally, this method achieves a high degree of integration between data-driven artificial intelligence and domain expertise, leveraging the powerful ability of reinforcement learning to handle complex nonlinear optimization problems while ensuring that the solution strictly adheres to basic scientific principles, providing a more solid and reliable technical foundation for intelligent control of safety-critical systems such as tunnel ventilation.

[0095] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0096] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A big data-driven method for predicting tunnel gas concentration and scheduling ventilation, characterized in that, Includes the following steps: S1. Based on the multi-source heterogeneous historical data of the tunnel, geological feature clustering is performed using an unsupervised learning algorithm to divide the tunnel into multiple geological prediction units; S2. For each geological prediction unit, extract multi-dimensional features containing temporal correlations, and use an ensemble learning framework to train a dedicated gas concentration prediction model to form a model library. S3. During tunnel excavation, real-time data on the working conditions of construction machinery and rock debris images at the tunnel face are collected and integrated, and the current geological category is dynamically identified through a pre-trained geological identification model. S4. Match the dynamically identified geological categories with the geological prediction units in the model library: If a match is found, the corresponding dedicated gas concentration prediction model is used as the effective prediction model. If the matching fails, it is determined that the area has entered an unforeseen geological anomaly zone. The emergency ventilation plan is then activated, and the online model learning process is started simultaneously. The online learning process is based on the real-time data stream newly generated in the current anomaly zone. Using the prediction model of the most adjacent geological unit as the base model, incremental learning and rapid fine-tuning are performed to generate a temporary prediction model for the anomaly zone as an effective prediction model. S5. Input real-time environmental data into an effective prediction model to obtain gas concentration prediction values ​​for future periods. Based on these prediction values, and with the goals of optimal ventilation efficiency and minimum energy consumption, generate differentiated ventilation scheduling instructions in real time through an optimization algorithm to drive the ventilation equipment in the corresponding area to execute.

2. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 1, characterized in that, In step S1, the geological feature clustering is performed using an unsupervised learning algorithm as follows: Based on historical data such as rock strength, groundwater status, fault structure information, and historical hazardous gas concentrations, a geological-gas feature vector is constructed. The Toeplitz Inverse Covariance-based Clustering (TICC) algorithm, which introduces mandatory link constraints, is used to cluster geological-gas feature vectors. Among these, it is necessary to link constraints based on expert knowledge construction of historical lithology logging data and gas emission patterns, which is used to guide the algorithm to classify samples with similar lithology and similar gas emission patterns into the same geological prediction unit.

3. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 1, characterized in that, In step S4, the online learning process of the model specifically includes: From the historical data repository, a portion of historical samples are extracted based on the similarity of data distribution with the current nearest geological unit to form a replay sample set; The replay sample set is mixed with the newly generated real-time data stream in the current anomaly region and used together as training data for incremental learning. During incremental learning and rapid fine-tuning, the replay sample set is weighted more heavily than the new sample set to constrain the model parameters from deviating excessively from the original knowledge during fine-tuning.

4. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 3, characterized in that, In step S4, after determining that an unforeseen geological anomaly zone has been entered, the following steps are also performed: Real-time fusion of rock debris image data from the working face, tunneling parameters of construction machinery, and information on the concentration of trace special gases collected by environmental sensors; Based on the fused information, a pre-trained deep learning model is used to calculate the dynamic safety risk score of the current working surface; When the safety risk score exceeds the preset threshold, a high risk is determined, and the level and response measures of the emergency ventilation plan are adjusted accordingly.

5. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 1, characterized in that, In step S5, with optimal ventilation efficiency and minimum energy consumption as multiple objectives, differentiated ventilation scheduling instructions are generated in real time through an optimization algorithm. This is specifically achieved through the following steps: S5.1 Construct a comprehensive evaluation index system that includes the average wind speed in the tunnel, the wind speed non-uniformity coefficient, the peak concentration of harmful gases, and the total power of the ventilation system; calculate the membership degree value of each individual index based on the fuzzy logic method; use the improved Seagull optimization algorithm to optimize the membership degree value of each index to obtain the equilibrium solution that makes the membership degree value of all indices optimal as a whole, and determine the required air volume of each branch in the ventilation network to meet the predicted gas concentration value based on the equilibrium solution. S5.

2. The tunnel ventilation network is abstracted into a graph structure consisting of nodes and branches, where nodes represent tunnel faces, junctions, or ventilation openings, and branches represent tunnel sections. The wind resistance characteristics of each branch are recorded. The air volume and air pressure of all fans in the network are used as optimization variables, the required air volume obtained in step S5.1 is used as a constraint, and the optimization objective is to minimize the total energy consumption of the fans. The ventilation network optimization model is constructed. The deep deterministic strategy gradient reinforcement learning algorithm is used to solve the ventilation network optimization model. Through continuous interaction between the agent and the ventilation network environment, the fan collaborative control strategy that minimizes the total energy consumption is learned, and ventilation scheduling instructions are generated.

6. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 1, characterized in that, In step S1, the fusion process of multi-source heterogeneous historical data also includes a data augmentation step, which specifically involves: A conditional generative adversarial network is used to generate synthetic samples by using geological category labels obtained by clustering through the Topplitz Inverse Covariance Clustering (TICC) algorithm as conditional input. Calculate the Mahalanobis distance between the synthetic sample and the cluster center of its geological category, compare the Mahalanobis distance with a preset threshold, and retain the synthetic samples whose Mahalanobis distance is not greater than the preset threshold to obtain the filtered synthetic samples. The selected synthetic samples were added to the geological-gas feature vector set to expand the training dataset used to train the dedicated gas concentration prediction model.

7. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 4, characterized in that, In the online learning process of the model in step S4, the specific process of incremental learning and rapid fine-tuning includes: establishing a dynamically updated memory to store representative historical samples; when performing incremental learning, selecting samples from the memory through an importance sampling strategy and mixing them with the real-time data stream.

8. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 5, characterized in that, In step S5.2, the solution process of the ventilation network optimization model specifically includes: modeling the fan cooperative control problem as a partially observable Markov decision process, where each fan is an agent, and using a multi-agent reinforcement learning algorithm for distributed training, sharing local observation information through the communication mechanism between agents.

9. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 4, characterized in that, The calculation process of dynamic safety risk score further includes: constructing a multimodal fusion module, extracting features from rock debris image data, tunneling parameters and trace special gas concentration information through convolutional neural network, fully connected layer and temporal encoder respectively, and using cross-modal attention mechanism to fuse these features.

10. The method for predicting tunnel gas concentration and scheduling ventilation based on big data as described in claim 5, characterized in that, In step S5.2, when constructing the ventilation network optimization model, physical constraints are further introduced. The Darcy-Weisbach equation and continuity equation in fluid mechanics are embedded as constraints into the reward function of the deep deterministic policy gradient reinforcement learning algorithm.