Data and algorithm double-driven tunnel construction global collaborative management system and method

Abstract

The application relates to the technical field of intelligent tunnel construction management, and discloses a data and algorithm double-driven global collaborative management system and method for tunnel construction, wherein the data and algorithm double-driven global collaborative management method for tunnel construction comprises the following steps: acquiring multi-source heterogeneous construction data and performing standardized pretreatment; extracting construction real-time state information and constructing a three-dimensional model, and dynamically binding the real-time state information and the three-dimensional model; constructing a process network model; performing real-time updating on the process network model; performing construction progress deduction; calculating a safety step distance and performing overrun judgment; extracting safety state information except the safety step distance, and performing risk analysis; identifying progress deviation; identifying risk response demand; and generating a regulation and control instruction; the application realizes process logic network modeling by fusing field knowledge constraints, progress deduction by deep reinforcement learning enhanced by migration learning, risk intelligent early warning by a time sequence prediction model, and global collaborative regulation and control by multi-target optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for tunnel construction, and more specifically, to a global collaborative management system and method for tunnel construction driven by both data and algorithms. Background Technology

[0002] Tunnel engineering, as a crucial component of transportation infrastructure, is characterized by complex and variable geological conditions, stringent requirements for process coordination, and numerous safety risks. With the rapid development of infrastructure construction in my country, the number of long tunnels and tunnel projects under complex geological conditions is increasing, placing higher demands on the precision and intelligence of construction management. Especially in scenarios involving parallel construction across multiple work faces, achieving precise control over the entire construction process, timely early warning of potential risks, and optimal allocation of limited resources have become core technical challenges that urgently need to be addressed in the field of tunnel engineering management.

[0003] In recent years, new-generation information technologies such as digital twins, industrial internet, and artificial intelligence have been widely applied in the field of engineering construction, providing technical support for intelligent management of tunnel construction. Existing solutions mainly focus on the collection and storage of construction information, visualization of 3D models, and threshold-based early warning for single indicators, achieving initial informatization of construction management. However, existing solutions still have significant shortcomings in real-time fusion of multi-source heterogeneous data, dynamic progress simulation based on process logic, proactive risk warning based on time-series prediction, and global resource optimization through multi-objective collaboration, making it difficult to meet the actual needs of global collaborative management in complex construction scenarios.

[0004] Specifically, existing technologies have the following technical problems: multi-source heterogeneous construction data are difficult to integrate in real time; there is a state deviation between digital twin models and physical entities, which cannot provide accurate situational awareness support for real-time decision-making; there is a lack of dynamic progress simulation capabilities based on process logic constraints, and when the actual progress deviates from the plan, it cannot automatically analyze the impact and simulate adjustment schemes; risk warning mainly relies on threshold judgment of current data, lacks the ability to predict the time series of risk evolution trends, and is difficult to achieve forward-looking identification of risks. Summary of the Invention

[0005] This invention provides a global collaborative management system and method for tunnel construction driven by both data and algorithms, which solves the technical problems of data fusion difficulties, lack of progress prediction, delayed risk warning, and insufficient collaborative control in related technologies.

[0006] This invention provides a data- and algorithm-driven global collaborative management method for tunnel construction, including: Acquire multi-source heterogeneous construction data, perform standardized preprocessing, and obtain a standardized construction data stream; Based on standardized construction data flow, real-time construction status information is extracted and a three-dimensional model is constructed. The real-time status information is dynamically bound to the three-dimensional model to obtain a living digital twin. Based on a living digital twin, a process network model is constructed; combined with standardized construction data flow, the process network model is updated in real time to obtain dynamically updated process parameters and process logical constraints. Based on the dynamically updated process parameters and process logic constraints, the construction progress is extrapolated to obtain dynamic progress prediction results and multi-scenario simulation results. Based on the live digital twin, the safe step distance is calculated and over-limit judgment is made to obtain the real-time over-limit warning information; based on the standardized construction data stream, the safety status information other than the safe step distance is extracted, risk analysis is performed, and graded risk warning information is obtained. Based on dynamic progress prediction results and multi-scenario simulation results, progress deviations are identified; based on over-limit real-time early warning information and graded risk early warning information, risk response needs are identified; and combined with progress deviations and risk response needs, control instructions are generated to obtain a global collaborative control plan and execution feedback data.

[0007] In a preferred embodiment, the step of acquiring multi-source heterogeneous construction data and performing standardized preprocessing includes: Edge computing nodes are deployed at each work site, and data acquisition adapters and local cache queues are configured to access and preprocess raw data from the nearest location to obtain raw construction data. Acquire raw construction data, perform preliminary data quality checks at edge nodes, and identify and mark abnormal data points; The cloud-based data platform receives data uploaded from various edge nodes and performs cross-data source timestamp alignment processing. Perform deep quality cleaning on the time-aligned data to remove outliers and fill in missing values; Perform data integrity checks and generate data quality reports to obtain standardized construction data streams.

[0008] In a preferred embodiment, the step of extracting real-time construction status information and constructing a three-dimensional model, and dynamically binding the real-time status information with the three-dimensional model, includes: Perform coordinate system unification processing on BIM and GIS models to establish spatial registration relationships between the two types of models; Construct a unified spatial index structure for the BIM-GIS fusion model to support spatial queries and correlation analysis across models; Define data-driven visualization mapping rules and establish the binding relationship between monitoring data and the visual attributes of model components; Establish a real-time data subscription and scenario incremental update mechanism; It provides multi-level scene browsing and interaction capabilities. The global overview mode displays the overall construction of the entire tunnel project, the working face focus mode displays the detailed status of a single working face, and the component details mode displays the attribute information of a single model component and the associated historical monitoring data curves.

[0009] In a preferred embodiment, the step of constructing a process network model based on a living digital twin and updating the process network model in real time by combining standardized construction data flow includes: Identify the process type, establish a process ontology model, and store the process ontology model in the form of a structured knowledge base to obtain the process ontology knowledge base; Based on the process ontology knowledge base, analyze the logical constraint relationships between processes and construct a process relationship edge set; A probabilistic prediction model for process duration is established based on the process ontology knowledge base to predict the duration distribution of each process. Based on the process relationship edge set, the hard constraints in the process specification are encoded into logical rules and embedded into the process network model; Establish an online learning and dynamic updating mechanism for process parameters.

[0010] In a preferred embodiment, the construction progress simulation includes: The construction progress prediction problem is formalized as a Markov decision process, defining the state space, action space and reward function; A deep Q-network is constructed as a decision model for progress prediction, and the network structure and training strategy are designed. The decision-making model for progress projection is optimized using transfer learning techniques; Embed process specification constraints into the deduction process; Using the current actual construction status as the initial state, the simulation model is run to generate progress predictions and multi-scenario simulation results.

[0011] In a preferred embodiment, calculating the safe step distance and performing an over-limit judgment includes: Extract the spatial location information of each working face and calculate the safe step distance in real time; Query the allowed safe step distance, perform an over-limit judgment, and generate an immediate warning; When the actual safe step distance exceeds the allowable threshold, an immediate over-limit warning message is generated. The warning message includes the working face number, the current safe step distance value, the allowable value, and the over-limit range.

[0012] In a preferred embodiment, the risk analysis includes: Construct a multi-dimensional risk time series prediction model based on long short-term memory networks; A sliding window approach is used to continuously run the risk prediction model and generate dynamic risk probability curves. Perform tiered early warning judgments and generate structured early warning information; Multiple warning thresholds are set for each type of risk, corresponding to different risk levels and response requirements.

[0013] In a preferred embodiment, the step of identifying progress deviations based on dynamic progress prediction results and multi-scenario simulation results, and identifying risk response requirements based on real-time over-limit warning information and graded risk warning information, includes: The dynamic progress forecast results are compared with the original construction plan to calculate the progress deviation of each work surface and each process. The graded risk warning information is transformed into the constraints of the resource optimization allocation model, and the real-time safety step value in the over-limit instant warning information is used as the quantitative parameter of the constraint. Construct a multi-objective resource optimization and allocation model, and define the optimization objectives and constraints.

[0014] In a preferred embodiment, generating control instructions by combining schedule deviations and risk response requirements includes: The optimization model is solved using a multi-objective evolutionary algorithm based on non-dominated sorting. Query the equipment capability profile and convert the optimization results into categorized control commands; Establish a closed-loop mechanism for instruction execution tracking and status feedback; For each control command issued, the system records its issuance time and expected execution time, and continuously tracks the execution status.

[0015] In a preferred embodiment, the data- and algorithm-driven global collaborative management system for tunnel construction is used to execute the aforementioned data- and algorithm-driven global collaborative management method for tunnel construction, including: The data acquisition module is used to acquire multi-source heterogeneous construction data, perform standardized preprocessing, and obtain a standardized construction data stream; The digital twin module extracts real-time construction status information based on standardized construction data flow and constructs a three-dimensional model. It dynamically binds the real-time status information with the three-dimensional model to obtain a living digital twin. The process modeling module constructs a process network model based on a living digital twin; combined with standardized construction data flow, the process network model is updated in real time to obtain dynamically updated process parameters and process logical constraints. The progress prediction module performs construction progress prediction based on dynamically updated process parameters and process logic constraints, and obtains dynamic progress prediction results and multi-scenario simulation results. The risk warning module, based on a live digital twin, calculates the safe step distance and judges whether it exceeds the limit, and obtains real-time warning information for exceeding the limit; it extracts safety status information other than the safe step distance based on standardized construction data stream, performs risk analysis, and obtains graded risk warning information. The collaborative control module identifies progress deviations based on dynamic progress prediction results and multi-scenario simulation results; identifies risk response needs based on real-time over-limit warning information and graded risk warning information; and generates control instructions by combining progress deviations and risk response needs to obtain a global collaborative control plan and execution feedback data.

[0016] The beneficial effects of this invention are as follows: by constructing an edge-cloud collaborative data acquisition architecture and a digital twin scenario that deeply integrates BIM and GIS, the real-time fusion of multi-source heterogeneous data at the construction site and the synchronous update of the living twin are realized. This solves the problems of untimely data acquisition, serious information silos, and disconnect between visualization and reality in traditional systems, and provides construction management with real-time situational awareness capabilities across all elements and all times, thereby improving the efficiency of managers in controlling the status of the construction site.

[0017] By integrating domain knowledge-constrained process logic network modeling, transfer learning-enhanced deep reinforcement learning progress extrapolation, time-series prediction model for intelligent risk early warning, and multi-objective optimization for global collaborative control, an intelligent closed-loop system of situational awareness, extrapolation and prediction, and dynamic control has been constructed. This has enabled a paradigm shift in tunnel construction management from passive response to proactive control, and from experience-driven to data and algorithm-driven. It has effectively solved technical problems such as lack of progress extrapolation capabilities, delayed risk early warning, and lack of collaborative control mechanisms, providing a systematic intelligent management solution for parallel construction of long tunnels with multiple working faces under complex geological conditions. Attached Figure Description

[0018] Figure 1 This is the main flowchart of the data and algorithm-driven global collaborative management method for tunnel construction in this invention; Figure 2 This is a detailed flowchart of the data- and algorithm-driven global collaborative management method for tunnel construction in this invention; Figure 3 This is a module diagram of the global collaborative management system for tunnel construction driven by both data and algorithms in this invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a data- and algorithm-driven global collaborative management method for tunnel construction, such as... Figures 1 to 2 As shown, it includes: Step 1: Obtain multi-source heterogeneous construction data, perform standardized preprocessing, and obtain a standardized construction data stream; Step 1.1, Edge computing node deployment and data access configuration; Edge computing nodes are deployed at each working face, configured with data acquisition adapters and local cache queues to enable local access and preprocessing of raw data. Based on the equipment distribution at the construction site, one edge computing node is set up at each working face. This node connects to various data sources within its working face area via wired or wireless means, including tunneling equipment control systems, surrounding rock deformation monitoring sensors, environmental monitoring instruments, personnel positioning base stations, material weighing systems, etc. The edge node has a built-in data acquisition adapter module that adapts and converts the communication protocols and data formats of different data sources, uniformly converting the raw data into an internal standard format to obtain uniformly formatted raw construction data. The edge node is also configured with a local cache queue. When the network connection is normal, data is uploaded in real time. When the network is interrupted, the data is temporarily stored in the local queue and re-uploaded in chronological order after the network is restored, ensuring the continuity and integrity of the data.

[0021] Step 1.2, Initial quality check and anomaly marking of edge data; Based on the standardized raw construction data output in step 1.1, the edge nodes perform a preliminary data quality check to identify and mark obviously abnormal data points, reducing the processing burden on the cloud. Upon receiving the raw data, the edge nodes perform a data integrity check, verifying whether the data packet structure conforms to the expected format and whether the fields are complete; secondly, they perform a data range check, comparing the values ​​of each monitoring indicator with preset reasonable ranges, marking data points that significantly exceed the reasonable range as suspected anomalies; and thirdly, they perform a data timeliness check, marking data points with timestamps significantly lagging behind the current time as delayed data. The data that has undergone the preliminary quality check, along with quality marking information, is uploaded to the cloud to obtain quality-marked construction data.

[0022] Step 1.3: Cloud data reception and timestamp alignment processing; Based on the construction data with quality tags output in step 1.2, the cloud-based data platform receives data uploaded from each edge node and performs cross-data source timestamp alignment. Because different data sources have different acquisition frequencies—the acquisition cycle for surrounding rock deformation monitoring data is typically on the order of minutes, while the acquisition cycle for tunneling equipment operating parameters can reach the order of seconds—it is necessary to align data of different frequencies to a unified time reference. The cloud-based data platform uses a linear interpolation method for time alignment. For data sources with lower acquisition frequencies, an estimated value is inserted between two adjacent actual acquisition points to unify the time resolution of all data sources to the standard period set by the system. During interpolation calculation, based on the values ​​and timestamps of two adjacent actual data points, the estimated value of the target time point is calculated according to a linear proportional relationship to obtain the time-aligned construction data.

[0023] Step 1.4, Deep data quality cleaning and missing value imputation; Based on the time-aligned construction data output in step 1.3, deep quality cleaning is performed on the time-aligned data to remove outliers and fill in missing values. The cloud data platform uses a sliding window-based statistical method for outlier detection. For each monitoring indicator, its mean and standard deviation within the sliding window are calculated. When the value of a data point deviates from the mean by more than a preset multiple of the standard deviation, it is identified as an outlier and removed. For missing values ​​caused by outlier removal or data transmission loss, a filling method based on historical data from the same period is used to fill in the missing values, i.e., data under the same working conditions in historical data are used as the reference for filling in the missing values, resulting in cleaned construction data.

[0024] The matching of working conditions employs a multi-dimensional similarity evaluation method, primarily including the following standards: The surrounding rock grade matching standard requires that the historical data and current data have completely consistent surrounding rock grades, or that they are adjacent grades in the surrounding rock classification system and have similar lithology. The cross-sectional size matching standard requires that the tunnel cross-sectional area difference does not exceed 10%, and that the cross-sectional shape type is the same, such as both being horseshoe-shaped or both being circular. The construction method matching standard requires that the excavation methods be consistent, such as both using drill-and-blast or both using shield tunneling, and that the support methods be the same, such as both using shotcrete and anchor support or both using segmental support. The geological condition matching standard includes a groundwater level difference not exceeding five meters, similar ground stress levels, and similar development levels of special geological structures such as fault fracture zones. The construction environment matching standard includes the same construction season or similar climatic conditions, and comparable technical levels and equipment configurations of the construction teams. The matching method employs weighted similarity calculation, assigning weights to each matching dimension. The weight for surrounding rock grade is the highest at 0.4, followed by cross-sectional dimensions at 0.25, construction method at 0.2, geological conditions at 0.1, and construction environment at 0.05. Similarity scores between the current and historical working conditions are calculated for each dimension, and then weighted summed to obtain the overall similarity score. Historical data with the highest similarity score exceeding 0.8 is selected as a reference.

[0025] Step 1.5: Data integrity verification and quality report generation; Based on the cleaned construction data output in step 1.4, data integrity verification is performed and a data quality report is generated. The cloud-based data platform periodically calculates the data coverage rate of each data source within a specified time range, which is the ratio of the actual number of valid data points received to the theoretically expected number. When the coverage rate of a data source falls below a preset threshold, the system automatically generates a data quality alarm, prompting relevant personnel to check the corresponding acquisition equipment or network link. The data quality report includes indicators such as coverage rate, anomaly rate, and latency rate for each data source, providing a basis for data governance. This results in a data quality report and a standardized construction data stream, which contains multi-source heterogeneous construction data that has undergone format standardization, quality checks, time alignment, anomaly cleansing, and integrity verification.

[0026] Furthermore, due to the harsh environment at tunnel construction sites, sensors may drift or fail due to factors such as vibration, dust, and humidity, making it difficult to accurately identify gradual anomalies using only statistical methods. Therefore, a deep learning-based anomaly detection method based on autoencoders can be employed to improve the ability to identify gradual anomalies and complex anomaly patterns. Specifically, an autoencoder neural network model is constructed, consisting of an encoder and a decoder. The encoder compresses the input multidimensional monitoring data into a low-dimensional latent vector representation, and the decoder restores the latent vectors to reconstructed data of the same dimension as the input. The autoencoder is trained using historical monitoring data under normal operating conditions, allowing it to learn the distribution characteristics of normal data. During online detection, real-time data is input into the trained autoencoder, and the reconstruction error between the input data and the reconstructed data is calculated. When the reconstruction error exceeds a preset threshold, it is considered an anomaly. This method can capture complex relationships between multidimensional data and has better detection performance for gradual anomalies and multivariate coupled anomalies.

[0027] Step 2: Extract real-time construction status information based on standardized construction data flow and construct a three-dimensional model. Dynamically bind the real-time status information with the three-dimensional model to obtain a living digital twin. Step 2.1, BIM-GIS model coordinate system unification and spatial registration; The coordinate systems of the BIM and GIS models are unified to establish spatial registration relationships between the two types of models. BIM models typically use a local project coordinate system, establishing a 3D rectangular coordinate system with the tunnel entrance or a specific control point as the origin; GIS models use a geographic coordinate system or a projected coordinate system, such as the WGS84 coordinate system or the Gauss-Kruger projection coordinate system. To achieve the fusion of the two types of models, coordinate transformation relationships need to be established.

[0028] Several control points are selected at the construction site. These control points have clear coordinate values ​​in both the BIM model and the GIS model. Based on the coordinate correspondence of the control points, the least squares method is used to solve for the coordinate transformation parameters, including translation parameters, rotation parameters, and scale parameters. The solved transformation parameters are then applied to transform the coordinates of all components in the BIM model to the GIS coordinate system, resulting in a BIM-GIS fusion model with unified coordinates.

[0029] Step 2.2, Unified Spatial Index Construction and Cross-Model Association Analysis; Based on the coordinate-unified BIM-GIS fusion model output in step 2.1, a unified spatial index structure for the BIM-GIS fusion model is constructed to support cross-model spatial queries and correlation analysis. The coordinate-transformed BIM model components and GIS model elements are uniformly incorporated into the spatial index management. Model data is organized using spatial index structures such as octrees or R-trees, supporting fast queries based on spatial extent. Octrees are suitable for handling evenly distributed point data and location queries in 3D space, while R-trees are more suitable for handling area or volume data with clear spatial boundaries and queries involving overlapping extents. Simultaneously, the topological relationships between BIM components and GIS elements are established, including the tunnel structure's crossing relationship with the surrounding rock strata, the intersection relationship between the tunnel entrance and the ground surface, and the fit relationship between the construction access road and the terrain. This provides a data foundation for subsequent spatial analysis, resulting in a spatially indexed fusion model.

[0030] Step 2.3, Define data-driven visualization mapping rules; Based on the standardized construction data stream output from Step 1 and the spatially indexed fusion model output from Step 2.2, data-driven visualization mapping rules are defined to establish the binding relationship between monitoring data and the visual attributes of model components. Corresponding visualization mapping rules are designed for different types of monitoring data. For surrounding rock deformation monitoring data, a color mapping method is used, mapping deformation values ​​to a gradient color spectrum from green to red, with larger deformation values ​​leaning more towards red. For construction progress data, a transparency mapping method is used, displaying completed sections as opaque and uncompleted sections as semi-transparent. For equipment status data, an icon labeling method is used, overlaying icons indicating the operating status at the equipment's location. All mapping rules are stored in configuration files, allowing for flexible adjustments based on user needs, resulting in a visualization mapping rule library.

[0031] Step 2.4, Real-time data subscription and scene incremental update mechanism; Based on the standardized construction data stream output in step 1 and the visualization mapping rule base output in step 2.3, a real-time data subscription and scene incremental update mechanism is established to achieve synchronization between the digital twin and the physical entity. The digital twin scene rendering engine subscribes to the standardized construction data stream output in step 1. When new data arrives, it triggers corresponding visualization update operations based on the data type and associated model components. To improve update efficiency, an incremental update strategy is adopted, re-rendering only the components whose state has changed, rather than the entire scene. Simultaneously, a data buffering mechanism is set up to merge and process multiple update messages arriving within a short period, avoiding performance issues caused by frequent rendering and obtaining a real-time updated 3D scene.

[0032] Step 2.5, Multi-level scene browsing and multi-granular situational awareness; Based on the real-time updated 3D scene output in step 2.4, multi-level scene browsing and interaction capabilities are provided, supporting multi-granular situational awareness from global to local perspectives. The digital twin scene supports three levels of browsing modes: the global overview mode displays the overall construction progress of the entire tunnel project, presenting the location distribution and overall progress of each working face from a bird's-eye view; the working face focus mode displays the detailed status of a single working face, presenting details such as the excavation face, support section, and equipment location from a first-person perspective; and the component details mode displays the attribute information of a single model component and the associated historical monitoring data curves. Users can switch between different levels through mouse or touch operations to achieve a comprehensive perception of the construction situation and obtain a living digital twin. This living digital twin includes a BIM-GIS fusion model with unified coordinates, a real-time data-driven visualization scene, and a multi-level interactive interface, which can realistically reflect the real-time status of the tunnel construction site.

[0033] Furthermore, since BIM models typically have higher geometric accuracy while GIS models have relatively lower geometric accuracy, direct fusion may lead to visual problems such as mismatched model boundaries. An adaptive model fusion method based on hierarchical detail technology can be adopted to optimize rendering performance while ensuring visual quality. Specifically, multi-level detail versions are constructed for both the BIM and GIS models, with each level corresponding to a different degree of geometric simplification. During scene rendering, the appropriate level of detail is dynamically selected based on the distance between the current viewpoint and the model. A high-detail version is used when the distance is close to ensure visual accuracy, while a low-detail version is used when the distance is far to reduce the rendering burden. Simultaneously, a transition zone is set at the boundary between the BIM and GIS models, and a gradual fusion method is used to handle the boundary connection between the two types of models, avoiding obvious seams or misalignments.

[0034] Step 3: Based on the live digital twin, construct a process network model; combine the standardized construction data flow to update the process network model in real time, and obtain dynamically updated process parameters and process logical constraint relationships; Step 3.1, Process type identification and ontology model establishment; Based on the construction progress and process information in the living digital twin output in step 2, the system systematically identifies the types of work processes and establishes a work process ontology model. It analyzes the types of work processes involved in the entire tunnel construction process, including but not limited to surveying and setting out, drilling, charging, blasting, ventilation and smoke extraction, muck removal and transportation, initial shotcrete, steel frame installation, wire mesh installation, secondary shotcrete, invert excavation, invert pouring, invert filling, waterproofing layer laying, secondary lining reinforcement binding, and secondary lining concrete pouring. An ontology definition is established for each work process type, including attributes such as work process name, work process code, construction stage, standard operating procedure, typical duration range, required resource type, and required workspace. The work process ontology model is stored in the form of a structured knowledge base, supporting query and reasoning operations, resulting in a work process ontology knowledge base.

[0035] Step 3.2, Analysis of process logic constraints and construction of edge sets; Based on the process ontology knowledge base output in step 3.1, the logical constraints between processes are analyzed, and a set of process relationship edges is constructed. The logical relationships between processes include the following types: sequential relationships, indicating that a process can only begin after another process is completed, such as initial shotcreting must be performed after slag removal; parallel relationships, indicating that multiple processes can be performed simultaneously, such as excavation operations on different working faces can be performed in parallel; conditional relationships, indicating that the initiation of a process requires meeting specific conditions, such as secondary lining needing to be performed after the initial support deformation has stabilized; and resource mutual exclusion relationships, indicating that multiple processes cannot be performed simultaneously when competing for the same resource, such as a lining trolley not serving two working faces at the same time. These logical relationships are encoded as process relationship edges, each edge containing attributes such as starting process, ending process, relationship type, and constraint parameters, resulting in a set of process relationship edges.

[0036] Step 3.3, Probabilistic prediction model for process duration based on historical data; Based on the process ontology knowledge base output in step 3.1 and historical construction data, a probabilistic prediction model for process duration is established. Actual duration data for various processes in historical projects under different surrounding rock grades, cross-sectional dimensions, and construction methods are collected, and statistical analysis methods are used to fit the probability distribution of the duration. For most processes, the duration distribution can be approximated by a beta distribution or a triangular distribution, with distribution parameters including minimum, maximum, and mode. When establishing the duration prediction model, factors such as surrounding rock grade, cross-sectional dimensions, and construction methods are used as input variables, and the duration distribution parameters are used as output variables. Regression analysis is used to establish the mapping relationship. This model can predict the duration distribution of each process based on the specific conditions of the current working face, providing a probabilistic duration estimate for schedule projection, thus obtaining the process duration probabilistic prediction model.

[0037] Step 3.4: Encoding of hard constraints in process specifications and embedding of logical rules; Based on the process relationship edge set output in step 3.2, the hard constraints in the process specifications are encoded into logical rules and embedded into the process logic network model. The construction process specifications contain several hard constraints that must be strictly followed, such as the safety step distance not exceeding the specified value, the concrete pouring process requiring a specified curing time before proceeding to the next process, and the use of short advances in special geological sections. These hard constraints are transformed into formal logical rules. The antecedents of the rules describe the conditions under which the constraints apply, and the consequents describe the requirements that must be met. These logical rules are associated with the process logic network model. In subsequent progress simulations, any simulation paths that violate the logical rules will be automatically eliminated, resulting in a constraint rule library.

[0038] Step 3.5, Online learning and dynamic updating mechanism for process parameters; Based on the process duration probability prediction model output in step 3.3 and the standardized construction data stream output in step 1, an online learning and dynamic update mechanism for process parameters is established. As the current project progresses, new actual process duration data is continuously accumulated. The system periodically incorporates the newly accumulated data into the training set of the duration prediction model, using incremental learning to update the model parameters, allowing the model to gradually adapt to the specific conditions of the current project. Incremental learning employs a sliding window strategy, prioritizing the use of recent data to update the model, enabling the model to capture dynamic changes in construction conditions. Simultaneously, a model performance monitoring mechanism is set up; when the prediction error exceeds a preset threshold, model retraining is triggered, resulting in dynamically updated process parameters and a process logic network model. This process logic network model integrates a process ontology knowledge base, process logical constraints, a process duration probability prediction model library, and a dynamic update mechanism. The process logical constraints include the process relationship edge set output in step 3.2 and the constraint rule library output in step 3.4.

[0039] Furthermore, since traditional process network models can only express binary relationships between processes, they are insufficient to describe complex constraints and implicit knowledge between multiple processes. Therefore, a knowledge graph-based process relationship modeling method can be adopted to enhance the model's ability to express complex process constraints. Specifically, the process logic network is extended into a knowledge graph structure. Nodes in the graph include not only process entities but also resource entities, spatial entities, temporal entities, constraint entities, and other types. Edges between nodes express not only sequential relationships but also occupancy relationships, location relationships, constraint relationships, and other semantics. Based on the knowledge graph representation, complex constraints can be expressed through graph query language, such as querying all processes occupying the same space within a certain time period or querying the availability status of a resource at a certain point in time. The knowledge graph also supports rule-based reasoning, which can automatically discover implicit process conflicts or resource bottlenecks.

[0040] Step 4: Based on the dynamically updated process parameters and process logic constraints, perform construction progress simulation to obtain dynamic progress prediction results and multi-scenario simulation results. Step 4.1, Markov decision process modeling for the schedule deduction problem; Based on the process logic network model output from step 3, the construction schedule deduction problem is formalized as a Markov decision process, defining the state space, action space, and reward function. The state space design needs to comprehensively describe the current state of the construction system, including the completion state vector of each process, the occupancy state vector of each resource, the current mileage of each working face, the cumulative used time, the current surrounding rock grade, and other variables. The action space design needs to cover all possible scheduling decisions, including starting a process in a ready state, pausing a process in progress, and transferring a resource from one working face to another. The reward function design needs to guide the deduction process towards the desired direction, setting positive rewards to encourage timely completion of processes and efficient resource utilization, setting negative rewards to penalize schedule delays and resource idleness, and setting large negative rewards to penalize decisions that violate process specification constraints, thus obtaining the Markov decision process model.

[0041] Step 4.2, Construction of Deep Q-Network Decision Model and Design of Training Strategy; Based on the Markov decision process model output in step 4.1, a Deep Q-Network (DQN) is constructed as the decision model for progress projection. The network structure and training strategy are designed. The Deep Q-Network takes the current state vector as input and outputs the Q-value estimate for each selectable action. The Q-value represents the expected cumulative reward after taking an action in the current state. The network structure adopts a multi-layer fully connected network. The input layer dimension is consistent with the state vector dimension, and the output layer dimension is consistent with the action space size. Several hidden layers are set in between for feature extraction and nonlinear transformation. The training process employs an experience replay technique, storing the state transition samples generated during projection in an experience pool. During training, batch samples are randomly sampled from the experience pool for gradient updates, breaking the temporal correlation between samples. Simultaneously, a target network technique is employed, setting a target network with lag parameter updates to calculate the target Q-value, improving training stability and resulting in the Deep Q-Network decision model.

[0042] Step 4.3: Optimize the inference model through transfer learning; Based on the deep Q-network decision model output in step 4.2, transfer learning technology is used to address the problem of insufficient training data for new projects. In the initial stage of system deployment, the current project has limited accumulated construction data, making it difficult to train a high-quality projection model from scratch. To solve this problem, construction data from multiple historical projects are collected, and the deep Q-network is pre-trained on this historical data, allowing the network to learn the general patterns of construction progress projection. The parameters of the pre-trained network are used as the initial parameters of the projection model for the current project, and the model is fine-tuned using the data already accumulated in the current project. A small learning rate is used during the fine-tuning process to avoid excessive coverage of pre-trained knowledge by new data. As the current project data continues to accumulate, the model gradually transitions from relying on historical knowledge to adapting to the characteristics of the current project, resulting in a projection model optimized through transfer learning.

[0043] Step 4.4: Feasibility assurance of embedding and extrapolating process specification constraints; Based on the transfer learning-optimized deduction model output from step 4.3 and the constraint rule base in the process logic network model output from step 3, process specification constraints are embedded into the deduction process to ensure the feasibility of the deduction results. In each deduction step, all theoretically possible actions are calculated based on the current state and the process logic network model. The constraint checking module is called to verify whether each candidate action violates the process specification constraints one by one. Actions that violate the constraints are removed from the candidate set, and only the action with the highest Q value is selected from the set of compliant actions for execution. This constraint embedding mechanism ensures that the deduction process always proceeds within the range allowed by the process specifications, and the generated schedule prediction results are practically feasible, resulting in a constraint-embedded deduction model.

[0044] Step 4.5: Schedule prediction and multi-scenario simulation result generation; Based on the constraint-embedded inference model output from step 4.4 and the current construction state in the living digital twin output from step 2, progress prediction and multi-scenario simulation results are generated. In progress prediction mode, the current actual construction state is used as the initial state, and the inference model is run until all processes are completed. The estimated start and finish times of each process are recorded during the inference process, generating a progress prediction Gantt chart and critical path analysis results, thus obtaining dynamic progress prediction results. In multi-scenario simulation mode, users can set different scenario assumption parameters, such as increasing or decreasing the quantity of a certain type of resource, assuming a section encounters adverse geological conditions, or assuming a certain equipment malfunctions. The system reruns the inference model under the modified conditions, generating progress simulation results under that scenario for decision-makers to compare and analyze, thus obtaining multi-scenario simulation results.

[0045] Furthermore, due to the efficiency bottleneck of deep Q-networks in handling continuous action spaces and high-dimensional state spaces, a deep reinforcement learning method based on the actor-critic framework can be adopted to improve the decision-making efficiency and quality of the inference model in complex scenarios. Specifically, a soft actor-critic algorithm is used to replace the deep Q-network, which includes two components: a policy network and a value network. The policy network directly outputs the probability distribution of actions, supporting the handling of continuous action spaces; the value network evaluates the value of the current state, providing gradient signals for updating the policy network. The soft actor-critic algorithm introduces a maximum entropy regularization term to encourage the policy to maintain a certain degree of exploratory nature and avoid getting trapped in local optima. This method can more efficiently search for the optimal scheduling strategy when dealing with complex scenarios such as parallel construction on multiple work surfaces.

[0046] Step 5: Based on the live digital twin, calculate the safe step distance and make an over-limit judgment to obtain an immediate over-limit warning; based on the standardized construction data stream, extract the safety status information other than the safe step distance, perform risk analysis, and obtain graded risk warning information. Step 5.1: Extraction of the spatial position of the working face and real-time calculation of the safety step distance; Based on the live digital twin output from step 2, the spatial location information of each working face is extracted, and the safe step distance is calculated in real time. The safe step distance refers to the distance between the excavation face and the completed initial support section, and is a core indicator for tunnel construction safety management. The system obtains the current mileage position of the excavation face and the mileage position of the completed initial support from the digital twin scenario; the difference between the two is the current safe step distance. The mileage position information comes from two channels: one is directly obtaining the excavation face position through the positioning system of the tunneling equipment; the other is indirectly calculating it through construction progress data, i.e., updating the excavation face position based on the most recent blasting advance record. The system continuously calculates the safe step distance at a set period and stores the calculation results in a time-series database for subsequent analysis, obtaining real-time safe step distance data.

[0047] Step 5.2: Query the allowable safe step distance and generate an over-limit warning; Based on the real-time safe step distance data output in step 5.1 and the surrounding rock grade information in the living digital twin output in step 2, the system queries the allowable safe step distance value, performs an over-limit judgment, and generates an immediate warning. Different surrounding rock grades correspond to different allowable safe step distance values; the worse the surrounding rock conditions, the smaller the allowable safe step distance. The system maintains a lookup table of surrounding rock grades and allowable safe step distance values, compiled according to tunnel construction technical specifications. After calculating the current safe step distance, the system queries the surrounding rock grade of the current working face segment, obtains the corresponding allowable value from the lookup table, and compares the actual safe step distance with the allowable value. When the actual value exceeds the allowable value, the system immediately generates an over-limit warning message, which includes the working face number, the current safe step distance value, the allowable value, and the over-limit range.

[0048] Step 5.3: Construct a multi-dimensional risk time series prediction model based on long short-term memory networks; Based on the standardized construction data stream output from Step 1, a multi-dimensional risk time series prediction model based on a Long Short-Term Memory Network (LSTM) is constructed. Besides the safety margin, tunnel construction also faces various risks such as surrounding rock instability, water inrush and mudslides, and excessive levels of hazardous gases. These risks typically exhibit certain precursory characteristics, manifested as specific trends in relevant monitoring indicators before their occurrence. The system constructs an LSTM network model, using historical sequences of multi-dimensional monitoring data as input, including observations of indicators such as surrounding rock deformation rate, convergence displacement, water inrush volume, and gas concentration over several past time steps. The output is a probability prediction of various risks occurring over several future time steps. The LSTM network selectively remembers and forgets historical information through a gating mechanism, effectively capturing long-term dependencies and trend characteristics in the monitoring data. Model training uses monitoring data from historical projects labeled with risk events, and supervised learning is used to optimize network parameters, resulting in the risk time series prediction model.

[0049] Step 5.4, Sliding window risk prediction and dynamic probability curve generation; Based on the risk time-series prediction model output in step 5.3 and the standardized construction data stream output in step 1, the risk prediction model is continuously run using a sliding window approach to generate a dynamic risk probability curve. The system triggers risk prediction calculations at set intervals. Each calculation extracts the monitoring data sequence from the most recent time window from the real-time database and inputs it into the trained Long Short-Term Memory (LSTM) network model to obtain the predicted risk probability values ​​for each future time step. The results of multiple consecutive predictions are plotted as a dynamic risk probability curve, with time on the horizontal axis and risk probability on the vertical axis. Different types of risks are represented by different colored curves. By observing the changing trend of the risk probability curve, managers can intuitively judge the evolution of the risk and obtain the dynamic risk probability curve.

[0050] Step 5.5: Risk classification and early warning judgment and generation of structured early warning information; Based on the dynamic risk probability curve output in step 5.4, a tiered early warning judgment is performed to generate structured early warning information. The system sets multiple early warning thresholds for each type of risk, corresponding to different risk levels and response requirements. When the predicted probability of a certain type of risk exceeds the first-level threshold, a blue warning is generated, indicating a potential risk requiring attention; when it exceeds the second-level threshold, a yellow warning is generated, indicating a higher risk requiring preventative measures; and when it exceeds the third-level threshold, a red warning is generated, indicating an urgent risk requiring immediate response. The early warning information uses a structured format, including fields such as warning number, warning type, warning level, warning location, warning time, risk probability value, and suggested response measures, facilitating subsequent automated processing and manual decision-making reference to obtain tiered risk early warning information.

[0051] Furthermore, since a single Long Short-Term Memory (LSTM) network model struggles to simultaneously capture both short-term fluctuations and long-term trends in monitoring data, a hybrid prediction model based on the fusion of temporal convolutional networks and LTM networks can be employed. The aim is to enhance the risk prediction's ability to capture features across different time scales. Specifically, a dual-branch network structure is constructed: one branch uses a temporal convolutional network to extract local patterns and short-term fluctuations from the monitoring data, while the other branch uses an LTM network to extract long-term dependencies and trend features. The output feature vectors of the two branches are adaptively fused using an attention mechanism, and the fused feature vector is input into a fully connected layer to generate the final risk probability prediction. The attention mechanism dynamically adjusts the contribution weights of the two branches based on the characteristics of the current input data, relying more on the output of the temporal convolutional network when short-term fluctuations are significant, and more on the output of the LTM network when long-term trends are evident.

[0052] Step 6: Based on the dynamic progress prediction results and multi-scenario simulation results output in Step 4, as well as the over-limit real-time early warning information and graded risk early warning information output in Step 5, obtain the control requirements, generate control instructions using multi-objective optimization and equipment capability adaptation methods, and obtain the global collaborative control scheme and execution feedback. Step 6.1: Identify and quantify schedule deviations and adjustment requirements; Based on the dynamic schedule prediction results and multi-scenario simulation results output in step 4, schedule deviations are identified and the effects of different adjustment strategies are analyzed. The dynamic schedule prediction results output in step 4 are compared with the original construction plan to calculate the schedule deviation for each work surface and each process. Schedule deviation is quantified using a deviation rate index, which is the difference between the actual or predicted duration and the planned duration divided by the planned duration. When the schedule deviation rate of a process exceeds a preset allowable threshold, the system marks that process as requiring schedule adjustment and determines the adjustment priority based on the direction and magnitude of the deviation. The resource allocation schemes in the multi-scenario simulation results output in step 4 are used as candidate adjustment strategies. By comparing the duration, resource utilization, and risk level under each scenario, feasible adjustment directions are selected. Processes lagging behind in schedule require increased resource input or optimized process connections to accelerate progress. Processes ahead of schedule can consider releasing some resources to support lagging processes, thus obtaining schedule adjustment needs and candidate adjustment strategies.

[0053] Step 6.2, Identification of risk response needs and establishment of constraints; Based on the over-limit real-time early warning information and graded risk early warning information output in step 5, risk response requirements are identified, and risk constraints are established. The graded risk early warning information output in step 5 is transformed into constraints for the resource optimization allocation model, while the real-time safety step distance value in the over-limit real-time early warning information is used as the quantitative parameter of the constraints. For over-limit real-time early warning information, the constraints require accelerating the support process of the corresponding working face or temporarily suspending the excavation process to bring the safety step distance back to the allowable range. The constraint strength is dynamically adjusted according to the over-limit amplitude in the over-limit real-time early warning information output in step 5.2. For surrounding rock instability risk early warning, the constraints require increasing the monitoring frequency and support strength at the early warning location, and suspending construction operations in the area if necessary. For water inrush risk early warning, the constraints require preparing drainage equipment and emergency supplies in advance. Risk constraints are incorporated into the optimization model in the form of hard constraints or soft constraints. Hard constraints must be strictly satisfied, while soft constraints guide the optimization direction through penalty terms, resulting in the risk constraints.

[0054] Step 6.3, Construction of a multi-objective resource optimization and allocation model; Based on the schedule adjustment requirements and candidate adjustment strategies output in step 6.1 and the risk constraints output in step 6.2, a multi-objective resource optimization allocation model is constructed, defining optimization objectives and constraints. The multi-scenario simulation results output in step 4 are used as the initial solution set and constraint reference for the optimization model. Resource allocation schemes performing well in the multi-scenario simulation results are used as the initial population for the evolutionary algorithm to improve convergence speed. The optimal project duration under each scenario serves as a reference benchmark for the optimization objective, avoiding the generation of unrealistic optimization objectives. The decision variables of the optimization model include the allocation quantity of various resources on each work surface, the planned start time of each process, and the equipment allocation scheme between work surfaces. The optimization objectives include three aspects: the first objective is to minimize the total project duration, i.e., minimizing the latest completion time of all processes; the second objective is to maximize resource utilization, i.e., maximizing the proportion of effective working time of resources to total available time; and the third objective is to minimize risk exposure, i.e., minimizing the number and duration of processes under risk. Constraints include process logic constraints, resource capacity constraints, spatial conflict constraints, and risk response constraints. Since there may be conflicts between the three optimization objectives, a multi-objective optimization method is used to solve the problem, resulting in the multi-objective optimization model.

[0055] Step 6.4: Solve using a multi-objective evolutionary algorithm based on non-dominated sorting; Based on the multi-objective optimization model output in step 6.3, a multi-objective evolutionary algorithm based on non-dominated sorting is used to solve the optimization model. The multi-objective evolutionary algorithm searches the solution space in a population-based manner, with each individual representing a feasible resource allocation scheme. During population initialization, candidate adjustment strategies output in step 6.1 are used as some initial individuals to improve the quality of the algorithm's initial solutions. The algorithm flow includes population initialization, fitness evaluation, non-dominated sorting, crowding calculation, selection, crossover, and mutation. Non-dominated sorting stratifies individuals in the population according to Pareto dominance, with the first layer being the Pareto front, containing all non-dominated solutions. Crowding calculation measures the distribution density of individuals in the objective space, used to maintain solution diversity when selecting within the same layer. After several generations of evolution, the algorithm outputs a set of non-dominated solutions on the Pareto front, each solution representing a resource allocation scheme that achieves different trade-offs among the three objectives. The system can automatically select one solution as the final scheme based on preset objective priorities, or it can present multiple solutions to the decision-maker for manual selection to obtain the optimal resource allocation scheme.

[0056] Step 6.5: Equipment capacity file query and type-based control command conversion; Based on the optimal resource allocation scheme output in step 6.4, the system queries the equipment capability profile and converts the optimization results into categorized control instructions. The system maintains an equipment capability profile, recording the intelligence level and supported control interfaces of each construction device. Intelligence levels are divided into three categories: Category 1 is fully automatic equipment, which supports receiving control instructions via the Industrial Internet and executing them automatically; Category 2 is semi-automatic equipment, which supports receiving work parameter settings but requires manual initiation; Category 3 is traditional equipment, which does not support remote control and requires complete manual operation. According to the equipment capability profile, the system converts the optimization results into different types of control instructions: For Category 1 equipment, it generates machine-executable control instructions, including fields such as equipment number, action type, action parameters, and execution time, and issues them via the Industrial Internet protocol; for Category 2 equipment, it generates parameter setting instructions, pushes them to the equipment operation terminal for operator confirmation and execution; for Category 3 equipment, it generates work guidance information, pushes it to the relevant team leader in text or chart form, and obtains categorized control instructions.

[0057] Step 6.6: Establish a closed-loop mechanism for instruction execution tracking and status feedback; Based on the categorized control commands output in step 6.5, a closed-loop mechanism for command execution tracking and status feedback is established. For each control command issued, the system records its issuance time and expected execution time, and continuously tracks the execution status. The execution status of the first type of equipment is automatically transmitted back via the Industrial Internet, including command reception confirmation, execution start, execution completion, and execution anomaly status. The execution status of the second and third types of equipment is fed back through manual confirmation; operators submit execution confirmation via mobile terminals after completing their tasks. The system summarizes the execution status of all commands, updates the corresponding status information in the live digital twin output in step 2, and completes the full closed loop of perception, deduction, control, and feedback. For commands that fail to provide feedback due to execution anomalies or timeouts, the system generates an anomaly alarm and triggers a manual intervention process, obtaining execution feedback and a global collaborative control plan.

[0058] Furthermore, since traditional multi-objective evolutionary algorithms converge slowly when dealing with large-scale optimization problems, a surrogate model-assisted multi-objective evolutionary algorithm can be adopted. The aim is to accelerate the optimization process while ensuring solution quality. Specifically, a Gaussian process regression model is constructed as a surrogate model for the objective function. The surrogate model can approximate the objective function value of candidate solutions with low computational cost. In each generation of the evolutionary algorithm, the surrogate model is used to quickly filter a large number of candidate solutions, selecting those with better predicted objective values. Only the selected candidate solutions are accurately evaluated using the true objective function. The results of the accurate evaluation are used to update the surrogate model, improving its prediction accuracy. This surrogate model-assisted strategy can reduce the number of evaluations of the true objective function, shortening the solution time to an acceptable range in multi-workface, large-scale resource optimization scenarios.

[0059] A data- and algorithm-driven global collaborative management system for tunnel construction, such as Figure 3 As shown, the data- and algorithm-driven global collaborative management method for tunnel construction, as described above, includes: The data acquisition module is used to acquire multi-source heterogeneous construction data, perform standardized preprocessing, and obtain a standardized construction data stream; The digital twin module extracts real-time construction status information based on standardized construction data flow and constructs a three-dimensional model. It dynamically binds the real-time status information with the three-dimensional model to obtain a living digital twin. The process modeling module constructs a process network model based on a living digital twin; combined with standardized construction data flow, the process network model is updated in real time to obtain dynamically updated process parameters and process logical constraints. The progress prediction module performs construction progress prediction based on dynamically updated process parameters and process logic constraints, and obtains dynamic progress prediction results and multi-scenario simulation results. The risk warning module, based on a live digital twin, calculates the safe step distance and judges whether it exceeds the limit, and obtains real-time warning information for exceeding the limit; it extracts safety status information other than the safe step distance based on standardized construction data stream, performs risk analysis, and obtains graded risk warning information. The collaborative control module identifies progress deviations based on dynamic progress prediction results and multi-scenario simulation results; identifies risk response needs based on real-time over-limit warning information and graded risk warning information; and generates control instructions by combining progress deviations and risk response needs to obtain a global collaborative control plan and execution feedback data.

[0060] In one embodiment of the present invention, a specific example is provided: A mountain tunnel project, approximately twelve kilometers long, was constructed using the drill-and-blast method, with multiple auxiliary tunnels forming an eight-face simultaneous excavation pattern. The tunnel traverses an area with complex geological conditions, including fault fracture zones and water-rich karst areas. Numerous intelligent devices and monitoring sensors were deployed at the construction site, enabling global collaborative management through the method described in this invention.

[0061] Table 1 shows an example of multi-source construction data collected by the system at a certain moment: Table 1: Example of multi-source construction data collected by the system at a certain moment; Table 2 shows examples of global coordinated control commands generated by the system: Table 2: Examples of global collaborative control commands generated by the system; By applying the method described in this invention, real-time situational awareness, dynamic progress prediction, intelligent risk warning, and global collaborative control of the entire tunnel construction process are realized, forming an intelligent closed-loop system of situational awareness, prediction, and dynamic control.

[0062] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A data- and algorithm-driven global collaborative management method for tunnel construction, characterized in that: include: Acquire multi-source heterogeneous construction data, perform standardized preprocessing, and obtain a standardized construction data stream; Based on standardized construction data flow, real-time construction status information is extracted and a three-dimensional model is constructed. The real-time status information is dynamically bound to the three-dimensional model to obtain a living digital twin. Based on a living digital twin, a process network model is constructed; combined with standardized construction data flow, the process network model is updated in real time to obtain dynamically updated process parameters and process logical constraints. Based on the dynamically updated process parameters and process logic constraints, the construction progress is extrapolated to obtain dynamic progress prediction results and multi-scenario simulation results. Based on the live digital twin, the safe step distance is calculated and over-limit judgment is made to obtain the real-time over-limit warning information; based on the standardized construction data stream, the safety status information other than the safe step distance is extracted, risk analysis is performed, and graded risk warning information is obtained. Identify schedule deviations based on dynamic schedule prediction results and multi-scenario simulation results; Identify risk response needs based on real-time over-limit early warning information and graded risk early warning information; By combining schedule deviations and risk response requirements, control instructions are generated, resulting in a global collaborative control plan and execution feedback data.

2. The data- and algorithm-driven global collaborative management method for tunnel construction as described in claim 1, characterized in that, The acquisition of multi-source heterogeneous construction data and the standardization preprocessing include: Edge computing nodes are deployed at each work site, and data acquisition adapters and local cache queues are configured to access and preprocess raw data from the nearest location to obtain raw construction data. Acquire raw construction data, perform preliminary data quality checks at edge nodes, and identify and mark abnormal data points; The cloud-based data platform receives data uploaded from various edge nodes and performs cross-data source timestamp alignment processing. Perform deep quality cleaning on the time-aligned data to remove outliers and fill in missing values; Perform data integrity verification and generate a data quality report to obtain a standardized construction data stream.

3. The data- and algorithm-driven global collaborative management method for tunnel construction according to claim 1, characterized in that, The step of extracting real-time construction status information and constructing a three-dimensional model, and dynamically binding the real-time status information with the three-dimensional model includes: Perform coordinate system unification processing on BIM and GIS models to establish spatial registration relationships between the two types of models; Construct a unified spatial index structure for the BIM-GIS fusion model to support spatial queries and correlation analysis across models; Define data-driven visualization mapping rules and establish the binding relationship between monitoring data and the visual attributes of model components; Establish a real-time data subscription and scenario incremental update mechanism; It provides multi-level scene browsing and interaction capabilities. The global overview mode displays the overall construction of the entire tunnel project, the working face focus mode displays the detailed status of a single working face, and the component details mode displays the attribute information of a single model component and the associated historical monitoring data curves.

4. The data- and algorithm-driven global collaborative management method for tunnel construction as described in claim 1, characterized in that, The construction of a process network model based on a living digital twin, and the real-time updating of the process network model in conjunction with standardized construction data flow, includes: Identify the process type, establish a process ontology model, and store the process ontology model in the form of a structured knowledge base to obtain the process ontology knowledge base; Based on the process ontology knowledge base, analyze the logical constraint relationships between processes and construct a process relationship edge set; A probabilistic prediction model for process duration is established based on the process ontology knowledge base to predict the duration distribution of each process. Based on the process relationship edge set, the hard constraints in the process specification are encoded into logical rules and embedded into the process network model; Establish an online learning and dynamic updating mechanism for process parameters.

5. The data- and algorithm-driven global collaborative management method for tunnel construction according to claim 1, characterized in that, The construction progress simulation includes: The construction progress prediction problem is formalized as a Markov decision process, defining the state space, action space and reward function; A deep Q-network is constructed as a decision model for progress prediction, and the network structure and training strategy are designed. The decision-making model for progress projection is optimized using transfer learning techniques; Embed process specification constraints into the deduction process; Using the current actual construction status as the initial state, the simulation model is run to generate progress predictions and multi-scenario simulation results.

6. The data- and algorithm-driven global collaborative management method for tunnel construction according to claim 1, characterized in that, The calculation of the safe step distance and the determination of exceeding the limit include: Extract the spatial location information of each working face and calculate the safe step distance in real time; Query the allowed safe step distance, perform an over-limit judgment, and generate an immediate warning; When the actual safe step distance exceeds the allowable threshold, an immediate over-limit warning message is generated. The warning message includes the working face number, the current safe step distance value, the allowable value, and the over-limit range.

7. The data- and algorithm-driven global collaborative management method for tunnel construction according to claim 1, characterized in that, The risk analysis includes: Construct a multi-dimensional risk time series prediction model based on long short-term memory networks; A sliding window approach is used to continuously run the risk prediction model and generate dynamic risk probability curves. Perform tiered early warning judgments and generate structured early warning information; Multiple warning thresholds are set for each type of risk, corresponding to different risk levels and response requirements.

8. The data- and algorithm-driven global collaborative management method for tunnel construction according to claim 1, characterized in that, The progress deviation is identified based on dynamic progress prediction results and multi-scenario simulation results; Risk response needs identified based on real-time over-limit early warning information and tiered risk early warning information include: The dynamic progress forecast results are compared with the original construction plan to calculate the progress deviation of each work surface and each process. The graded risk warning information is transformed into the constraints of the resource optimization allocation model, and the real-time safety step value in the over-limit instant warning information is used as the quantitative parameter of the constraint. Construct a multi-objective resource optimization and allocation model, and define the optimization objectives and constraints.

9. The data- and algorithm-driven global collaborative management method for tunnel construction according to claim 1, characterized in that, The generation of control instructions by combining schedule deviations and risk response requirements includes: The optimization model is solved using a multi-objective evolutionary algorithm based on non-dominated sorting. Query the equipment capability profile and convert the optimization results into categorized control commands; Establish a closed-loop mechanism for instruction execution tracking and status feedback; For each control command issued, the system records its issuance time and expected execution time, and continuously tracks the execution status.

10. A globally collaborative management system for tunnel construction driven by both data and algorithms, characterized in that: The method for global collaborative management of tunnel construction driven by both data and algorithms as described in any one of claims 1-9 includes: The data acquisition module is used to acquire multi-source heterogeneous construction data, perform standardized preprocessing, and obtain a standardized construction data stream; The digital twin module extracts real-time construction status information based on standardized construction data flow and constructs a three-dimensional model. It dynamically binds the real-time status information with the three-dimensional model to obtain a living digital twin. The process modeling module constructs a process network model based on a living digital twin; combined with standardized construction data flow, the process network model is updated in real time to obtain dynamically updated process parameters and process logical constraints. The progress prediction module performs construction progress prediction based on dynamically updated process parameters and process logic constraints, and obtains dynamic progress prediction results and multi-scenario simulation results. The risk warning module, based on a live digital twin, calculates the safe step distance and judges whether it exceeds the limit, and obtains real-time warning information for exceeding the limit; it extracts safety status information other than the safe step distance based on standardized construction data stream, performs risk analysis, and obtains graded risk warning information. The collaborative control module identifies progress deviations based on dynamic progress prediction results and multi-scenario simulation results; identifies risk response needs based on real-time over-limit warning information and graded risk warning information; and generates control instructions by combining progress deviations and risk response needs to obtain a global collaborative control plan and execution feedback data.

Images