Tunnel scheduling command system and method based on digital twinning and distributed decision

Abstract

The invention discloses a tunnel scheduling command system and method based on digital twinning and distributed decision, and belongs to the technical field of tunnel construction man-vehicle scheduling, and the system comprises a digital twinning engine module which is used for constructing a three-dimensional digital twinning model of a tunnel space; the data collaborative governance module is used for carrying out association analysis on historical scheduling decision data recorded in a distributed account book and converting an analysis result into a material scheduling instruction through an intelligent contract; the fusion positioning system is used for uploading positioning data of personnel and equipment in the tunnel to the edge computing cluster in real time; the meeting algorithm module is used for acquiring real-time traffic flow data and a prediction avoidance threshold value from the digital twin engine module, generating a decentralized passing strategy by combining adjacent vehicle positioning information provided by a fusion positioning system, and sending the decentralized passing strategy to the digital twin engine module; and the strategy execution result is issued to the transportation equipment after being subjected to authority verification through an intelligent contract of the data collaborative governance module. According to the invention, the safety and efficiency of tunnel scheduling are improved.

Description

Technical Field

[0001] This application relates to a tunnel scheduling and command system and method based on digital twins and distributed decision-making, belonging to the field of tunnel construction personnel and vehicle scheduling technology. Background Technology

[0002] In tunnel scheduling and management, traditional scheduling and command systems often struggle to meet the dynamic scheduling needs of complex scenarios due to issues such as data silos, delayed decision-making, and insufficient positioning accuracy. These challenges include delayed early warnings of personnel gathering risks, inefficient response to equipment failures, and frequent traffic flow conflicts when multiple vehicles meet. While existing digital twin technology can achieve virtual space mapping, it is limited by insufficient multi-source heterogeneous data fusion capabilities, limited edge computing resources, and a lack of distributed collaborative decision-making mechanisms, making it difficult to simultaneously achieve real-time and accurate scheduling strategies. Therefore, a tunnel scheduling and command system is urgently needed to achieve closed-loop management of the entire process, enabling accurate risk prediction, dynamic and flexible scheduling, and real-time efficient optimization. Summary of the Invention

[0003] According to one aspect of this application, a tunnel scheduling and command system based on digital twins and distributed decision-making is provided, which improves the safety and efficiency of tunnel scheduling.

[0004] A tunnel scheduling and command system based on digital twins and distributed decision-making, characterized in that it includes: The digital twin engine module is used to construct a three-dimensional digital twin model of the tunnel space. The three-dimensional digital twin model integrates BIM, GIS, multi-source environmental sensor data and dynamic entity mapping, outputs personnel gathering trends, equipment failure risks and traffic flow dynamics, and dynamically generates optimized scheduling strategies based on the prediction results. The data collaborative governance module is used to receive the predicted data output by the digital twin engine module, perform correlation analysis with the historical scheduling decision data recorded in the distributed ledger, convert the analysis results into material scheduling instructions through smart contracts, and feed them back to the digital twin engine module to update the scheduling strategy. The integrated positioning system is used to upload UWB positioning data of personnel and equipment in the tunnel to the edge computing cluster in real time through 5G network slicing, receive the positioning parsing results processed by the edge computing cluster, and synchronize them to the digital twin engine module to correct the dynamic entity mapping. The vehicle meeting algorithm module is used to obtain real-time traffic flow data and predicted avoidance thresholds from the digital twin engine module, combine them with the adjacent vehicle positioning information provided by the fusion positioning system, generate a decentralized traffic strategy, and send the strategy execution result to the transportation equipment after authorization verification through the smart contract of the data collaborative governance module. Edge computing clusters are used for localized processing of raw positioning data from fusion positioning systems. They output positioning analysis results through an AI inference engine, accelerate inference of machine learning models issued by the digital twin engine module, and send the results back to the vehicle meeting algorithm module to optimize decision-making timeliness.

[0005] Furthermore, the digital twin engine module includes: A multi-source data fusion unit is used to synchronously update environmental sensor data, equipment status data, and personnel positioning data within the tunnel to the three-dimensional digital twin model; The prediction model training unit trains a machine learning model based on historical data. The output of the machine learning model includes a population density heatmap, equipment failure probability, and traffic congestion warning.

[0006] Furthermore, the data collaborative governance module includes: The blockchain sub-module uses distributed ledger technology to store scheduling instruction execution logs, material flow records, and permission change information; The smart contract submodule has a preset scheduling rule library, which automatically performs resource allocation, permission verification, or task issuance when preset conditions are triggered.

[0007] Furthermore, the fusion positioning system includes: The 5G network slicing unit is configured to prioritize low latency in the transmission of scheduling instructions; UWB positioning units enable the location of personnel and equipment through UWB base stations deployed inside tunnels; Among them, based on the predefined signal coverage area of ​​the tunnel map, the signal strength is monitored through real-time sensor data, and the 5G signal is automatically switched to UWB positioning mode when it is lower than the preset threshold.

[0008] Furthermore, the meeting algorithm module includes: The reinforcement learning training unit uses vehicle traffic efficiency, safe distance and energy consumption as optimization goals to iteratively update the avoidance strategy; The dynamic priority allocation unit adjusts the passage priority in real time based on vehicle load, direction of travel, urgency of the task, tunnel congestion, and distance to the passing point.

[0009] Furthermore, the edge computing cluster includes: A lightweight AI inference engine, deployed on edge nodes at the tunnel site, supporting TensorFlow Lite or ONNXRuntime frameworks; The data encryption unit encrypts sensitive data using the national cryptographic SM4 algorithm before uploading it to the cloud.

[0010] Furthermore, the execution of the vehicle meeting algorithm module includes: Real-time acquisition of the operating status data of the current vehicle and surrounding vehicles, and receipt of global tunnel status information, together constructing a real-time environmental status vector; Based on the perception distance threshold dynamically calculated by the reinforcement learning model, it is determined whether to trigger the distributed collaborative decision-making process; The current vehicle, acting as an autonomous intelligent agent, inputs the environmental state vector into a locally deployed reinforcement learning policy network, outputs preliminary passage instructions, and calculates the real-time priority scores of the current vehicle and conflicting vehicles based on predefined weight rules. The vehicle broadcasts instructions and scores via vehicle-to-vehicle (V2V) communication, and in the event of instruction conflicts, arbitrates the vehicle to give way based on the real-time priority score. The vehicle that is giving way performs an evasive maneuver, and the vehicle with priority passes through the conflict area according to instructions. For parked vehicles, the startup operation is performed when the policy network predicts that startup will not cause a conflict, or when a system release instruction is received.

[0011] Furthermore, the reinforcement learning policy network is trained iteratively using vehicle traffic efficiency, safe distance, and energy consumption as multi-objective optimization functions; Furthermore, the calculation factors for the real-time priority score include at least: vehicle load, driving direction, task urgency, distance from the passing point, and their contribution to the overall congestion status of the tunnel.

[0012] Furthermore, the traffic efficiency includes the average travel time of vehicles from the tunnel entrance to the tunnel face, the number of vehicles passing through the key sections of the tunnel per unit time, and the average waiting time and avoidance facility utilization rate of vehicles at the conflict point. The safe distance includes the vehicle dynamic safe distance, the safe margin buffer zone, and the human-vehicle interaction safe distance; The energy consumption includes driving energy consumption, total fuel consumption, and energy costs.

[0013] According to another aspect of this application, a tunnel scheduling and command method based on digital twins and distributed decision-making is also proposed, including: Construct a three-dimensional digital twin model of the tunnel space, integrate BIM, GIS, multi-source environmental sensor data and dynamic entity mapping, and integrate machine learning models to predict personnel gathering trends and equipment failure risks in real time, and dynamically generate optimized scheduling strategies. Blockchain distributed ledger records scheduling decisions and material flow data to ensure that the data is tamper-proof and auditable throughout the entire process; The integrated positioning system is used to achieve positioning and low-latency communication in the tunnel. The signal adaptive switching mechanism is used to automatically switch to UWB standalone mode in the 5G signal coverage blind spot. The adaptive learning vehicle meeting algorithm is implemented, which enables the vehicle agent to optimize the traffic strategy in real time through a reinforcement learning model, and adjusts the avoidance threshold and priority based on historical traffic data and real-time environmental perception. It leverages edge computing clusters to process real-time data locally, accelerates the execution of machine learning models through an AI inference engine, and collaborates with a blockchain module to ensure data security.

[0014] The beneficial effects that this application can produce include: The tunnel scheduling and command system and method based on digital twins and distributed decision-making provided in this application constructs a tunnel digital twin model by integrating high-precision 3D modeling technology, achieves efficient fusion and intelligent analysis of multi-source heterogeneous data through multi-modal data collaborative governance, ensures the flexibility and reliability of scheduling strategies through a distributed intelligent decision-making mechanism, and accelerates real-time data processing and model inference with the help of edge computing. This not only significantly improves the prediction accuracy and response speed of risks such as personnel gathering, equipment failure, and traffic conflicts in tunnel scheduling, but also dynamically generates and corrects optimized scheduling strategies in real time. It effectively solves the problems of data silos, decision lag, and inaccurate positioning in traditional systems, ultimately achieving closed-loop management of the entire process of risk prediction, dynamic scheduling, and real-time optimization, greatly improving the safety and efficiency of tunnel scheduling. Attached Figure Description

[0015] Figure 1 This is a system block diagram of a tunnel scheduling and command system based on digital twins and distributed decision-making in one embodiment of this application; Figure 2 This is a flowchart of a tunnel scheduling and command method based on digital twins and distributed decision-making in one embodiment of this application. Detailed Implementation

[0016] The present application is described in detail below with reference to the embodiments, but the present application is not limited to these embodiments.

[0017] Example 1: See Figure 1 A tunnel scheduling and command system based on digital twins and distributed decision-making includes: The digital twin engine module is used to construct a three-dimensional digital twin model of the tunnel space. The three-dimensional digital twin model integrates BIM, GIS, multi-source environmental sensor data and dynamic entity mapping, outputs personnel gathering trends, equipment failure risks and traffic flow dynamics, and dynamically generates optimized scheduling strategies based on the prediction results. The data collaborative governance module is used to receive the predicted data output by the digital twin engine module, perform correlation analysis with the historical scheduling decision data recorded in the distributed ledger, convert the analysis results into material scheduling instructions through smart contracts, and feed them back to the digital twin engine module to update the scheduling strategy. The integrated positioning system is used to upload UWB positioning data of personnel and equipment in the tunnel to the edge computing cluster in real time through 5G network slicing, receive the positioning parsing results processed by the edge computing cluster, and synchronize them to the digital twin engine module to correct the dynamic entity mapping. The vehicle meeting algorithm module is used to obtain real-time traffic flow data and predicted avoidance thresholds from the digital twin engine module, combine them with the adjacent vehicle positioning information provided by the fusion positioning system, generate a decentralized traffic strategy, and send the strategy execution result to the transportation equipment after authorization verification through the smart contract of the data collaborative governance module. Edge computing clusters are used for localized processing of raw positioning data from fusion positioning systems. They output positioning analysis results through an AI inference engine, accelerate inference of machine learning models issued by the digital twin engine module, and send the results back to the vehicle meeting algorithm module to optimize decision-making timeliness.

[0018] Specifically, the digital twin engine module constructs a three-dimensional digital twin model that maps physical and digital spaces based on the tunnel's entire lifecycle data. The model integrates static structural data from BIM (Building Information Modeling), such as tunnel cross-sections, lining structures, and equipment installation locations; spatial topological data from GIS (Geographic Information System), such as tunnel alignment, entrance / exit coordinates, and surrounding terrain relationships; and dynamic real-time data from multi-source environmental sensors, such as temperature, humidity, visibility, harmful gas concentration, smoke concentration, and road surface smoothness. Simultaneously, through dynamic entity mapping technology, the model correlates in real-time with moving entities within the tunnel, including the location, status, and movement trajectory data of personnel, construction equipment, transport vehicles, and emergency rescue equipment, forming a comprehensive digital view encompassing static structure, dynamic entities, and environmental perception. Based on fused data, the module achieves three major predictions through built-in machine learning algorithms: First, by combining personnel location trajectories and work area density thresholds, it identifies the risk of illegal gatherings and predicts personnel gathering trends. Second, based on equipment operating parameters and historical fault data, it provides early warnings of potential problems such as mechanical wear and circuit failures, enabling equipment failure risk prediction. Third, by combining vehicle speed, traffic flow, and collision point data, it predicts congestion or collision risks, achieving dynamic traffic flow prediction. Based on the prediction results, the system dynamically generates multi-scenario optimized scheduling strategies, such as personnel diversion strategies for construction areas, priority maintenance scheduling instructions for faulty equipment, and staggered vehicle traffic route planning. These strategies can be dynamically iterated and corrected based on real-time feedback from the physical space, such as sensor data updates and changes in personnel / equipment positions, achieving closed-loop optimization of prediction, decision-making, execution, and feedback.

[0019] The data collaborative governance module receives predictive data from the digital twin engine, such as risk warnings and optimized scheduling suggestions. It also performs correlation analysis with historical scheduling decision data stored in the distributed ledger (built using blockchain technology), such as past fault handling records, traffic management plans, and material allocation ledgers. This data comparison verifies the rationality and feasibility of the scheduling strategy; for example, it optimizes the current maintenance scheduling timeline by referencing the historical handling efficiency of similar equipment faults. Based on the correlation analysis results, the module uses pre-set smart contracts (programmable automated execution protocols) to convert abstract scheduling strategies into standardized, executable material scheduling instructions. These instructions include: the scheduling object, such as construction vehicles and emergency material warehouses in a certain area; the task to be performed, such as going to a designated area for maintenance and transferring emergency materials to the fault location; time requirements, such as arriving within 30 minutes; and permission levels, such as priority and execution entity permission verification rules. After the material dispatching instructions are generated, they are sent to the corresponding execution terminals, such as equipment control consoles, personnel handheld terminals, and vehicle navigation systems. At the same time, they are simultaneously fed back to the digital twin engine module to update the execution status of the dispatching strategy, ensuring that the execution progress of the digital twin model and the physical space are synchronized in real time. Meanwhile, the entire dispatching process data is encrypted and written into the distributed ledger to form an immutable dispatching log, providing historical data support for subsequent strategy optimization.

[0020] The integrated positioning system focuses on high-precision, low-latency positioning of personnel and equipment within tunnels. It employs UWB positioning technology and 5G network slicing, using UWB positioning base stations deployed within the tunnel to collect raw positioning data of personnel and equipment in real time, such as distance, angle, and signal strength. Leveraging 5G network slicing technology, a dedicated high-speed channel is established for positioning data transmission, ensuring that raw data is uploaded to the edge computing cluster without delay or packet loss, avoiding the latency issues of traditional network transmission. This is particularly suitable for real-time transmission of positioning data in long-distance tunnels. The system receives the positioning analysis results processed by the edge computing cluster, such as the precise coordinates, movement speed, direction angle, and location of personnel / equipment, and synchronizes them to the digital twin engine module via a high-speed data interface. This is used to correct the accuracy of dynamic entity mapping, for example, eliminating positioning drift errors and ensuring that the position of personnel / equipment in the digital model is completely consistent with the physical space, providing accurate entity location data support for the prediction and decision-making of the digital twin engine.

[0021] The vehicle-passing algorithm module addresses traffic flow safety issues within tunnels, particularly in scenarios involving mixed traffic of construction vehicles and emergency vehicles. This module employs a decentralized decision-making model to avoid the delay risks associated with traditional centralized scheduling. It acquires real-time traffic flow data from a digital twin engine, including vehicle location, speed, traffic volume, and direction of travel. Based on preset safe passing distance and speed thresholds for tunnel width, curve radius, and vehicle dimensions, it predicts avoidance thresholds. Simultaneously, it combines high-precision positioning information (coordinates, relative distance, and trajectory prediction) of adjacent vehicles, especially those near potential conflict points, provided by a fusion positioning system, to construct a vehicle-passing conflict risk assessment model. Based on the risk assessment results, the module generates decentralized traffic strategies. These strategies do not rely on centralized commands from a central server; adjacent vehicles can collaboratively make decisions through edge nodes. Examples include: slowing down and yielding to vehicles on the outer side when passing on curves, prioritizing passage for vehicles on the inner side, allowing empty vehicles to yield to heavily loaded vehicles when passing on narrow roads, and planning for vehicles to move to the side of the road to prioritize emergency vehicle passage. After the strategy is generated, it needs to be verified for permissions through the smart contract of the data collaboration governance module, such as confirming the scheduling permissions of the vehicle and whether the strategy complies with the tunnel traffic rules. After verification, it is directly sent to the vehicle terminal of the transportation equipment to ensure that the strategy is legal, compliant and accurately executed. At the same time, the execution results, such as whether the vehicle avoids according to the strategy, the time spent on passing other vehicles, and whether there is a conflict, are sent back to the digital twin engine and the distributed ledger for subsequent algorithm optimization.

[0022] To address the pain points of high data transmission latency and heavy load on central servers in tunnel scenarios, this edge computing cluster module adopts an edge computing architecture, deploying computing power at edge nodes at tunnel entrances / exits or in the middle area to achieve localized data processing and localized decision-making. It receives raw UWB positioning data uploaded by the fused positioning system and performs real-time analysis using built-in positioning algorithms, such as TOF (Time of Flight) or TDOA (Time Difference of Arrival), outputting positioning results to avoid latency caused by uploading raw data to a remote cloud. It also performs localized accelerated inference on machine learning models from the digital twin engine, such as risk prediction models and vehicle collision assessment models, utilizing GPU / TPU computing power clusters to improve model computation efficiency. The processed positioning analysis results and model inference results are synchronously transmitted back to the vehicle collision algorithm module and the digital twin engine module to ensure the timeliness and accuracy of scheduling decisions. By leveraging localized computing power, the data transmission distance and latency are significantly reduced, addressing the core pain points of slow data transmission and delayed decision response in long and complex tunnel scenarios. At the same time, it reduces the computing load on the central server, improves the overall stability and anti-interference capabilities of the system, and allows edge nodes to independently complete basic positioning and decision processing even in the event of local fluctuations in the 5G network.

[0023] The digital twin engine module includes: A multi-source data fusion unit is used to synchronously update environmental sensor data, equipment status data, and personnel positioning data within the tunnel to the three-dimensional digital twin model; The prediction model training unit trains a machine learning model based on historical data. The output of the machine learning model includes a population density heatmap, equipment failure probability, and traffic congestion warning.

[0024] Specifically, the multi-source data fusion unit is responsible for the efficient access, cleaning, and integration of multi-dimensional and multi-type data, ensuring the timeliness and accuracy of the model data. Through standardized data interfaces, such as the MQTT IoT communication protocol and HTTP / HTTPS interfaces, it synchronizes tunnel environmental sensor data, equipment status data, and personnel positioning data to the 3D digital twin model in real time. The tunnel environmental sensor data includes real-time monitoring data collected by temperature and humidity sensors, visibility detectors, hazardous gas sensors, smoke sensors, and road surface condition sensors; the data update frequency is configurable according to scenario requirements. Equipment status data includes operating parameters of fixed equipment within the tunnel, such as ventilation fans, lighting systems, and fire-fighting equipment, as well as operating data of mobile devices, such as construction machinery and transport vehicles, uploaded in real time through equipment controllers or vehicle terminals. Personnel positioning data receives high-precision coordinates of personnel synchronized from the fusion positioning system, combines them with personnel identification, and achieves a mapping between personnel identity, real-time location, and movement trajectory; the data update frequency is no less than once per second, ensuring real-time tracking of personnel dynamics.The raw data from synchronous access is preprocessed to eliminate data noise, fill in data gaps, and standardize and convert it according to the data format requirements of the 3D digital twin model. This includes associating sensor data with the spatial coordinates of corresponding monitoring points in the model and binding equipment status data to the attribute fields of equipment entities in the model. Ultimately, this achieves deep integration of environmental, equipment, and personnel data with the model, ensuring that the digital space accurately reflects the real-time status of the physical space. This application also provides an implementation scheme for real-time access to monitoring data of the tunnel physical space, including environmental, equipment, and personnel positioning data. The monitoring data is preprocessed and standardized, and then integrated into the 3D digital twin model. Specifically, this includes the following steps: Step S1: Dividing the continuous acquisition process into multiple time sub-stages; Step S2: Comparing the data fluctuation trend of each sub-stage with the data fluctuation trend of the entire acquisition stage; Step S3: If the trends are inconsistent, marking the sub-stage as a fluctuation stage and recording the trend. The process involves analyzing the interval between the initial moment and the end moment of the floating phase; if the trend is consistent, the sub-phase is marked as a stable phase, and the numerical increase span of the monitoring data in multiple consecutive stable phases is calculated. The data floating trend is quantified and compared using statistical methods or trend fitting algorithms; Step S4: Determine whether the interval duration exceeds the interval duration threshold or whether the numerical increase span exceeds the span threshold; Step S5: If the determination result of Step S4 is yes, it is determined that there is a lag anomaly, and the collection and prediction cycle of this type of monitoring data is shortened; wherein, floating real-time prediction is performed in the floating phase, a numerical floating span threshold is set in the stable phase, and real-time demand prediction is triggered when the monitoring data exceeds the threshold. Shortening the collection and prediction cycle specifically includes at least one of increasing the data sampling frequency, shortening the time window of the prediction model, or increasing the frequency of rolling prediction; Step S6: If the determination result of Step S4 is no, the original collection and prediction cycle is maintained. Therefore, this application divides the data collection phase of each dimension of environment, equipment, and personnel into several sub-phases. For example, environmental data is divided according to different time periods of the day; equipment data can be divided according to equipment operating cycles or specific task phases; personnel data can be divided according to personnel activity areas within the tunnel or task execution phases. The data collection process includes analyzing the fluctuation trends of dimensional data in each sub-stage and the overall fluctuation trends of dimensional data throughout the entire collection period. For environmental data, the analysis covers the changing trends of parameters such as temperature and humidity within each sub-stage, as well as the overall changing trends throughout the day. For equipment data, the analysis covers the changing trends of parameters such as fan speed and lighting brightness within each sub-stage, comparing them with the changing trends within the overall equipment operating cycle. For personnel data, the analysis covers the changing trends of personnel positions in specific areas within each sub-stage, as well as the overall trend of activity within the entire tunnel. If the corresponding fluctuation trends are inconsistent, the sub-stage is marked as a fluctuating stage; if the corresponding fluctuation trends are consistent, the sub-stage is marked as a stable stage.For example, in the environmental dimension, if the temperature rise trend in a certain sub-stage does not match the overall temperature rise trend throughout the day, that sub-stage is marked as a floating stage; if they match, it is marked as a stable stage. Within a floating stage, the interval between the time of data trend analysis and the end of the floating stage is recorded. For example, in the equipment dimension, if a ventilator's operating parameters exhibit abnormal fluctuations in a certain sub-stage, it is marked as a floating stage, and the time interval between the time of analyzing this abnormal trend and the end of that sub-stage is recorded. When the time interval within a floating stage exceeds a pre-set threshold, an anomaly in the dimensional data acquisition process is inferred. In this case, the service demand assessment platform shortens the dimensional data acquisition prediction cycle, i.e., performs real-time floating prediction within the floating stage. For example, for harmful gas concentration data in the environmental dimension, if the data update interval exceeds the threshold during a floating stage, the subsequent data acquisition prediction cycle is immediately shortened to monitor changes in harmful gas concentration in real time, allowing for timely implementation of ventilation and other measures. Within a stable stage, the range of values ​​collected from multiple stable stages during dimensional data trend analysis is increased. For example, in the personnel dimension, the analysis examines the range of changes in personnel location coordinates across multiple stable activity phases, i.e., the distance personnel move between adjacent stable phases. If the increase in value within a stable phase exceeds a pre-set threshold, an anomaly in the dimensional data acquisition process is inferred. The service demand assessment platform shortens the dimensional data acquisition prediction cycle and sets a critical value for the range of numerical fluctuations within stable phases; exceeding this threshold triggers timely demand forecasting. For instance, in the equipment dimension, for a lighting system, if the increase in lighting brightness values ​​across multiple sub-phases exceeds a threshold during a stable operation phase, it indicates a potential equipment malfunction or control anomaly. The data acquisition prediction cycle is immediately shortened, and a new brightness fluctuation threshold is set; exceeding this threshold triggers timely equipment maintenance demand forecasting. If the time interval within the fluctuating phase does not exceed the interval duration threshold, and the increase in value within the stable phase does not exceed the range threshold, the dimensional data acquisition process is considered to be normal. In this case, data acquisition and processing continue according to the original data update frequency and prediction cycle, ensuring the accurate reflection of the real-time state of the physical space by the digital space.

[0025] The predictive model training unit is responsible for training, optimizing, and iterating the machine learning model, providing algorithmic support for the risk prediction function of the digital twin engine, and ensuring the accuracy and reliability of the prediction results. It collects historical tunnel operation data from distributed ledgers and system databases, specifically including: personnel-related historical data (location trajectory data of past personnel operations, records of personnel gathering events, personnel evacuation drill data, etc.); equipment-related historical data (operating parameters throughout the equipment's lifecycle, fault records, such as fault type, fault occurrence time, pre-fault operating conditions data, maintenance plans and effects, equipment maintenance logs, etc.); and traffic flow-related historical data (traffic flow data at different times, vehicle trajectory data, records of oncoming traffic conflicts, traffic congestion event data, etc.). The collected historical data is labeled, such as labeling equipment operation data with "normal," "abnormal," and "fault," and personnel location data with "safe area," "dangerous area," and "gathering area," to construct a high-quality model training dataset. Based on labeled historical datasets, various machine learning algorithms are used to construct and train prediction models. The core trained models and their outputs include: a personnel density heatmap prediction model: using convolutional neural networks or density clustering algorithms, taking historical personnel location data, work area area, and personnel flow change patterns as input, the trained model can predict the personnel density distribution in various areas of the tunnel within the next 5-30 minutes based on real-time personnel location data, and visually outputting it as a heatmap. For example, red areas represent high-density clustering risk, yellow areas represent normal density, and green areas represent low density, assisting in identifying potential personnel gathering hazards; and an equipment failure probability prediction model: using gradient boosting trees or long short-term memory networks, taking historical equipment operating parameters, failure records, and other data as input. Using the maintenance cycle as input, the trained model can analyze the current equipment operating data in real time, output the probability values ​​of various equipment failures in the next 1-24 hours, and mark the high-risk failure types and possible time nodes, providing a basis for preventive maintenance of equipment; Traffic congestion early warning model: Using time series prediction algorithms or graph neural networks, with historical traffic flow data, tunnel terrain parameters (such as the location of curves and slopes), and vehicle type distribution as input, the trained model can predict the possibility of traffic congestion on various road sections in the next 10-60 minutes based on real-time traffic flow and vehicle speed data, such as "high risk", "medium risk" and "low risk", and output the road sections where congestion may occur and the expected congestion duration, providing early warning support for the formulation of traffic flow management strategies.The trained model is periodically iterated and updated based on newly generated operational data. The model's predictions are compared with actual occurrences, such as comparing the predicted probability of equipment failure with the actual occurrence of failures, and comparing the traffic congestion warning results with actual congestion events. Evaluation indicators such as model accuracy and recall are calculated. If the indicators are lower than the preset threshold, such as the accuracy being lower than 90%, the model parameters are readjusted, such as optimizing the algorithm hyperparameters and expanding the training dataset, to ensure that the model can adapt to changes in tunnel operation scenarios, such as adjustments to construction techniques, equipment updates, and changes in traffic flow patterns, and to maintain high prediction accuracy.

[0026] It is worth noting that, based on the real-time fused data provided by the multi-source data fusion unit and the machine learning model trained by the prediction model training unit, the digital twin engine module realizes prediction of personnel gathering trends, equipment failure risk, and traffic flow dynamics. Specifically, the output of the personnel density heatmap prediction model directly supports the judgment of personnel gathering trends, the results of the equipment failure probability prediction model provide a quantitative basis for equipment failure risk assessment, and the early warning information from the traffic congestion warning model assists in the dynamic analysis of traffic flow. Based on the prediction results, the system dynamically generates multi-scenario optimized scheduling strategies, including personnel diversion strategies for construction areas, priority maintenance scheduling instructions for faulty equipment, and staggered vehicle traffic route planning. These strategies can be dynamically iterated and corrected through real-time feedback data received by the multi-source data fusion unit (such as sensor data updates and changes in personnel / equipment positions). Simultaneously, new strategy execution data, such as personnel diversion effects, equipment maintenance results, and traffic management efficiency, are recorded and fed back to the prediction model training unit as historical data for the next model iteration.

[0027] The data collaborative governance module includes: The blockchain sub-module uses distributed ledger technology to store scheduling instruction execution logs, material flow records, and permission change information; The smart contract submodule has a preset scheduling rule library, which automatically performs resource allocation, permission verification, or task issuance when preset conditions are triggered.

[0028] Specifically, the blockchain subsystem leverages distributed ledger technology to construct a decentralized and tamper-proof data storage system, providing secure and reliable storage and traceability guarantees for the entire tunnel scheduling process. Core data during tunnel scheduling is encrypted to ensure data integrity and security. Detailed lifecycle information for each scheduling instruction is recorded, including instruction generation time, instruction content, instruction recipient, execution progress, execution result feedback, and exception handling records. Each log entry is uniquely timestamped and digitally signed to ensure immutability. Real-time tracking of the entire flow of various materials within the tunnel is also provided, covering material name, specifications, quantity, initial storage location, transfer time, recipient, transfer route, current storage location, and usage status. This enables full lifecycle traceability of materials from procurement and warehousing to consumption, preventing loss, misuse, or waste. Furthermore, the system records all user permission settings and changes, including user identification, initial permission scope, permission change application time, reason for change, approving body, changed permission scope, and effective time, ensuring compliance and traceability of permission management and preventing unauthorized operations and permission abuse. A multi-node distributed storage architecture is adopted, synchronously storing the aforementioned core data across multiple edge nodes and a central server node within the system, avoiding the risk of data loss due to the failure of a single node. Simultaneously, asymmetric encryption algorithms are used to encrypt the data, ensuring security during storage and transmission. Furthermore, once data is written to the distributed ledger, it cannot be unilaterally modified or deleted; any data changes require multi-node consensus verification, further ensuring data trustworthiness and immutability. Convenient data traceability and auditing functions are provided. Users can quickly query the complete flow record and storage node information of corresponding data by inputting key search conditions, such as instruction number, material code, user ID, and time range, forming a visualized traceability chain. The system can also automatically generate data audit reports, statistically analyzing key indicators such as scheduling instruction execution efficiency, material turnover rate, and frequency of permission changes, providing data support for tunnel operation management optimization and compliance auditing.

[0029] The smart contract submodule is based on a pre-set scheduling rule library to build a smart contract system that can be automatically triggered and executed. This enables the automation and standardization of material allocation, permission verification, and task issuance, reducing manual intervention and improving scheduling efficiency and accuracy. Combined with the actual needs of tunnel operation, such as daily construction scheduling, emergency rescue scheduling, and equipment maintenance scheduling, a standardized scheduling rule library covering multiple scenarios is built. The rule content includes triggering conditions, execution actions, execution subjects, permission requirements, and exception handling mechanisms. For example: Daily construction material dispatching rules: The trigger condition is "the remaining amount of materials in construction area A is lower than a preset threshold (e.g., the remaining amount of concrete is < 5 cubic meters)", the action is "allocate 5 cubic meters of concrete from material warehouse B to construction area A", the executing entity is "material transportation team", the permission requirement is "the transportation team must have access to construction area A", and the exception handling mechanism is "if the transportation vehicle malfunctions, automatically switch to the backup transportation vehicle and notify the dispatch administrator"; Equipment failure emergency dispatching rules: The trigger condition is "the probability of failure of digital twin engine module output device C is ≥ 90% (high-risk failure)", the action is "issue an emergency repair task for device C to the equipment maintenance team, and lock the operation right of device C at the same time". The system has several restrictions: 1) **Limited Access:** The implementing entity is the "Equipment Maintenance Team," and the access requirement is that "maintenance personnel must have the authority to operate and repair Equipment C." The exception handling mechanism is that "if the maintenance team's response delay exceeds 15 minutes, the emergency command center will be automatically notified." 2) **Dynamic Adjustment Rules for Personnel Access:** The trigger condition is "operator D enters a high-risk area (such as a tunnel blasting operation area)," and the execution action is "temporarily grant operator D access to the high-risk area, but automatically revoke the access after leaving the area." The implementing entity is the "System Access Management Module," and the access requirement is that "operator D must possess basic operational qualifications." The exception handling mechanism is that "if access granting fails, an audible and visual alarm will be automatically issued, and the safety administrator will be notified." The system also supports dynamic updates and maintenance of scheduling rules. Administrators can modify, add, or delete trigger conditions and execution actions in the rule base according to changes in tunnel operation scenarios. Modified rules must be verified through multi-node consensus before taking effect.The smart contract submodule receives real-time predictive data from the digital twin engine module, location data from the fusion positioning system, and historical data from the blockchain submodule. When the data meets the trigger conditions of the preset scheduling rules, it automatically executes corresponding actions: When the inventory of materials in a certain area is found to be below a threshold, the smart contract automatically generates a material allocation instruction, allocates materials from the nearest warehouse with sufficient inventory, and synchronizes the allocation instruction to the transportation equipment terminal and the blockchain submodule; Before the scheduling instruction is issued or personnel / equipment enter a specific area, the smart contract automatically verifies whether the permissions of the executing entity meet the requirements, such as whether the transportation vehicle has access to the target area and whether the maintenance personnel have equipment maintenance permissions. If the permissions meet the requirements, the instruction is allowed to be executed or the area is entered; if the permissions are insufficient, the operation is automatically rejected and the reason for the insufficient permissions is reported, and the permission verification log is recorded to the blockchain submodule; When equipment failure warnings or personnel gathering risks are detected, the smart contract automatically generates corresponding task instructions and issues them to the corresponding executing entity terminal according to the task type and priority, while setting the task execution time limit and monitoring the execution progress in real time. If the task is not completed within the time limit, the exception handling mechanism is automatically triggered. After executing the corresponding action, the smart contract submodule receives the execution result from the executing entity in real time and compares the result with the preset target. If the execution result meets the preset target, it is recorded in the blockchain submodule and simultaneously fed back to the digital twin engine module for updating the scheduling strategy; if the execution result does not meet the preset target, the exception handling mechanism is automatically triggered, and the exception is recorded in detail in the blockchain submodule.

[0030] It's worth noting that during operation, the data collaboration and governance module first receives predictions from the digital twin engine module. Then, the blockchain sub-module retrieves historical scheduling decision data stored in the distributed ledger and performs correlation analysis with the prediction data to verify the rationality of the scheduling strategy. After analysis, the smart contract sub-module, based on the scheduling rule base, converts the reasonable scheduling strategy into standardized scheduling instructions and automatically verifies the executor's permissions. Once approved, these instructions are sent to the execution terminal. During instruction execution, the blockchain sub-module records scheduling instruction execution logs, material flow records, and permission change information in real time, ensuring full traceability. The smart contract sub-module monitors the instruction execution progress in real time and handles execution anomalies. Finally, the instruction execution results are fed back to the digital twin engine module to update the 3D digital twin model and scheduling strategy. Simultaneously, new execution data is written to the blockchain sub-module, providing historical data support for subsequent scheduling decisions, forming a complete closed loop of data collaboration and instruction flow.

[0031] The fusion positioning system includes: The 5G network slicing unit is configured to prioritize low latency in the transmission of scheduling instructions; UWB positioning units enable the location of personnel and equipment through UWB base stations deployed inside tunnels; Among them, based on the predefined signal coverage area of ​​the tunnel map, the signal strength is monitored through real-time sensor data, and the 5G signal is automatically switched to UWB positioning mode when it is lower than the preset threshold.

[0032] Specifically, based on tunnel positioning and scheduling requirements, dedicated network slices are applied for from operators or self-built in a 5G standalone network environment. These slices are configured with independent bandwidth, computing, and routing resources, and are physically isolated from other service slices to avoid transmission delays caused by network congestion. Simultaneously, slice parameters are optimized for tunnel scenarios to control network latency within 10-20 milliseconds, meeting the low-latency transmission requirements for positioning data and scheduling commands, such as vehicle avoidance commands and emergency evacuation commands. A tiered transmission mechanism is established within dedicated data slices, prioritizing data and instructions based on their importance: Emergency dispatch instructions, such as emergency evacuation orders and equipment repair orders, and high-risk area location data, such as personnel location data in blasting operation areas, utilize the highest transmission resources within the slice to ensure zero-latency transmission; Regular dispatch instructions, such as material allocation orders and equipment inspection orders, and ordinary work area location data, such as daily location data of construction personnel, are transmitted efficiently during the intervals between highest-priority data transmissions, with latency controlled within 15 milliseconds; Non-real-time data, such as historical location trajectory backups and location data statistical reports, are transmitted during idle periods and do not consume core transmission resources. Priority management ensures that critical data and instructions are transmitted first, preventing data congestion from affecting the timeliness of dispatch decisions. The system monitors the transmission status of dedicated slices in real time, including metrics such as bandwidth utilization, transmission latency, and packet loss rate. Through built-in network optimization algorithms, it automatically requests temporary capacity expansion when bandwidth utilization exceeds 80%, automatically switches to backup routes when transmission latency exceeds a preset threshold, and initiates a data retransmission mechanism when packet loss rate exceeds 1% to ensure the stability and integrity of location data and command transmission.

[0033] UWB positioning units achieve high-precision location awareness by deploying UWB (Ultra-Wideband) positioning base stations within tunnels and combining this with UWB positioning tags worn / installed by personnel and equipment. Based on parameters such as tunnel length, cross-sectional structure, and turning radius, a deployment plan for UWB positioning base stations is developed. In straight tunnel sections, positioning base stations are deployed at appropriate intervals according to actual construction needs, while ensuring overlapping signal coverage areas between adjacent base stations to avoid positioning blind spots. In tunnel curves, positioning base stations are deployed on both sides of the curve to reduce signal attenuation caused by curve obstruction and ensure positioning accuracy in the curve area. In high-risk areas, such as blasting operation areas and areas with dense equipment, positioning base stations are deployed more densely to improve positioning accuracy in these areas. Simultaneously, based on the tunnel map, the signal coverage area of ​​each base station is predefined to form a visualized positioning coverage map, marking the positioning accuracy range of each area and providing a basis for coverage range analysis during subsequent signal switching.

[0034] UWB positioning tags are worn on personnel's wrists, safety helmets, or installed on equipment. They communicate with three or more nearby positioning base stations via ultra-wideband pulse signals. After receiving the signal from the tag, the base stations record the Time of Arrival (TOA) or Time Difference of Arrival (TDOA) and upload the raw time data to the edge computing cluster. The edge computing cluster uses the UWB positioning algorithm, combined with the known coordinates of each base station, to calculate the precise coordinates of the positioning tag, including X, Y, and Z three-dimensional coordinates. The Z coordinate is used to distinguish the height position of personnel / equipment in the tunnel cross-section, such as the equipment installation height or the personnel's standing position, thus achieving positioning. The working status of the UWB positioning tag is monitored in real time, including battery level, signal transmission frequency, and whether it is offline. When the tag's battery level is detected to be below 20%, a low battery reminder is automatically sent to the corresponding personnel or equipment management terminal. When the tag does not transmit a signal for 30 consecutive seconds, it is determined to be offline, automatically triggering an anomaly warning and notifying the dispatch administrator to investigate tag malfunctions or personnel / equipment abnormalities.

[0035] To address the issue of positioning interruptions caused by 5G signal obstruction and interference within tunnels, a dynamic signal switching mechanism has been established to ensure continuous and stable acquisition of positioning data. Signal monitoring sensors are installed at key locations within the tunnel to collect the signal strength of the dedicated 5G slices in real time. Simultaneously, the UWB positioning unit monitors the signal coverage strength of the positioning base station in real time. The monitored 5G signal strength is compared with a preset threshold to determine whether the 5G signal meets transmission requirements. The system automatically triggers signal switching when any of the following conditions are met: ① The 5G signal strength is below the preset threshold for 10 consecutive seconds and cannot be restored through dynamic bandwidth adjustment or routing switching; ② The 5G signal packet loss rate exceeds 5% for 5 consecutive seconds, causing positioning data transmission interruption; ③ Specific areas within the tunnel, such as the middle section of long tunnels or deeply buried tunnel sections, are designated as areas with weak 5G signal coverage. When personnel / equipment enter these areas, switching is directly triggered. Upon triggering the switchover, the system immediately ceases transmitting location data via 5G network slicing and automatically activates the UWB positioning unit. It then collects location data for personnel / equipment through UWB positioning base stations. The edge computing cluster rapidly parses the UWB positioning data, generates precise coordinates, and converts the switched location data format to a standard format consistent with 5G transmission data. This ensures seamless reception and processing by the digital twin engine module, preventing data gaps. The system sends a signal switchover notification to the dispatch administrator, informing them of the switchover reason, area, and time, allowing the administrator to monitor the positioning system status in real time. When the 5G signal strength is detected to be above a preset threshold for 20 consecutive seconds and the transmission status is stable, the system automatically switches back to 5G network slicing transmission mode, restoring the advantages of low-latency data transmission.

[0036] It is worth noting that during operation, the fusion positioning system uploads the raw positioning data collected by the UWB positioning unit to the edge computing cluster with low latency through the 5G network slicing unit. After parsing the data, the edge computing cluster generates positioning results and synchronizes them to the digital twin engine module to correct the dynamic entity mapping in the 3D digital twin model. At the same time, the positioning data is synchronized to the vehicle meeting algorithm module to provide data support for obtaining positioning information of adjacent vehicles and generating decentralized traffic strategies. When the system triggers a signal switch, the fusion positioning system feeds back the switch status and positioning data connection status to the data collaborative governance module. The data collaborative governance module adjusts the transmission path of the scheduling command through the smart contract sub-module and writes the signal switch record to the blockchain sub-module to ensure the traceability of the positioning system's operating status.

[0037] The vehicle meeting algorithm module includes: The reinforcement learning training unit uses vehicle traffic efficiency, safe distance and energy consumption as optimization goals to iteratively update the avoidance strategy; The dynamic priority allocation unit adjusts the passage priority in real time based on vehicle load, direction of travel, urgency of the task, tunnel congestion, and distance to the passing point.

[0038] The execution of the meeting algorithm module includes: Real-time acquisition of the operating status data of the current vehicle and surrounding vehicles, and receipt of global tunnel status information, together constructing a real-time environmental status vector; Based on the perception distance threshold dynamically calculated by the reinforcement learning model, it is determined whether to trigger the distributed collaborative decision-making process; The current vehicle, acting as an autonomous intelligent agent, inputs the environmental state vector into a locally deployed reinforcement learning policy network, outputs preliminary passage instructions, and calculates the real-time priority scores of the current vehicle and conflicting vehicles based on predefined weight rules. The vehicle broadcasts instructions and scores via vehicle-to-vehicle (V2V) communication, and in the event of instruction conflicts, arbitrates the vehicle to give way based on the real-time priority score. The vehicle that is giving way performs an evasive maneuver, and the vehicle with priority passes through the conflict area according to instructions. For parked vehicles, the startup operation is performed when the policy network predicts that startup will not cause a conflict, or when a system release instruction is received.

[0039] The reinforcement learning policy network is trained iteratively using vehicle traffic efficiency, safe distance and energy consumption as multi-objective optimization functions. Furthermore, the calculation factors for the real-time priority score include at least: vehicle load, driving direction, task urgency, distance from the passing point, and their contribution to the overall congestion status of the tunnel.

[0040] The traffic efficiency includes the average travel time of vehicles from the tunnel entrance to the tunnel face, the number of vehicles passing through the key sections of the tunnel per unit time, the average waiting time of vehicles at the conflict point, and the utilization rate of the avoidance facilities. The safe distance includes the vehicle dynamic safe distance, the safe margin buffer zone, and the human-vehicle interaction safe distance; The energy consumption includes driving energy consumption, total fuel consumption, and energy costs.

[0041] Specifically, the reinforcement learning training unit aims to improve vehicle traffic efficiency, ensure safe distances, and reduce energy consumption within the tunnel. It continuously iterates and optimizes oncoming traffic avoidance strategies using reinforcement learning algorithms, ensuring the strategies adapt to different traffic flow scenarios and tunnel environment changes. Traffic efficiency is quantified from three dimensions: first, the average travel time of vehicles from the tunnel entrance to the tunnel face (shorter travel time equals higher efficiency); second, the number of vehicles passing through key tunnel sections per unit time, such as sections with frequent oncoming traffic or the middle section of the tunnel (higher number of vehicles equals higher efficiency); and third, the average waiting time and utilization rate of avoidance facilities at conflict points (shorter waiting time and higher facility utilization rate equal higher efficiency). By collecting these three types of data in real time, a comprehensive traffic efficiency score is calculated; a higher score corresponds to a higher reward value in the optimization function. Safety distance encompasses three levels: first, dynamic vehicle safety distance, dynamically adjusted based on vehicle speed and road surface friction coefficient; for example, at 20 km / h, the dynamic safety distance is ≥4 meters; at 30 km / h, it is ≥6 meters. Second, safety margin buffer zone, an additional safety space reserved on top of the dynamic safety distance; for example, a buffer zone of ≥2 meters in tunnel curves and ≥1 meter in straight areas. Third, pedestrian-vehicle interaction safety distance, with a minimum safe distance of ≥5 meters between vehicles and personnel working in the tunnel. If the actual vehicle distance meets all safety distance requirements, the highest reward is given; if any safety distance fails to meet the standard, a tiered penalty is triggered, with greater deviations resulting in stronger penalties. Energy consumption is statistically analyzed from three dimensions: first, vehicle driving energy consumption, calculated based on engine power, driving speed, and load; for example, a truck with 80% load capacity traveling at 20 km / h consumes approximately 15 L / 100km; second, total fuel consumption (the total fuel consumed during the entire vehicle-to-vehicle encounter process); and third, energy cost (the economic cost calculated based on fuel consumption and fuel unit price). The model optimizes vehicle acceleration and deceleration frequencies, reducing sudden braking and acceleration, and lowering three types of energy consumption indicators. The lower the energy consumption, the higher the reward value in the optimization function. During training, an adaptive learning rate adjustment mechanism is adopted. When the model's achievement rate of optimization targets in a certain scenario, such as a congestion scenario (e.g., traffic efficiency ≥ 90%, safe distance compliance ≥ 95%), exceeds a preset threshold for three consecutive days, the learning rate is reduced to minimize parameter fluctuations. When the achievement rate falls below the threshold, the learning rate is increased to accelerate model iteration and ensure that the policy network can continuously adapt to changes in tunnel operation scenarios.

[0042] The dynamic priority allocation unit assigns passing priority to each vehicle in real time based on its own attributes, task urgency, and tunnel traffic congestion, ensuring efficient passage for critical vehicles while balancing overall traffic flow. Specifically, it adjusts passing priority in real time based on vehicle load, direction of travel, task urgency, tunnel congestion, and distance to a passing point. Among the vehicle attributes, the load factor includes, but is not limited to, the vehicle's actual load and maximum load (heavy load ≥ 80%, medium load 30%-80%, light load < 30%), with a load coefficient correction. For example, the coefficient for a fully loaded truck is 1.2, and for an empty truck it is 0.8. The base score = load level base score × load coefficient. Therefore, a heavy load base score is 8 points × 1.2 = 9.6 points, a medium load is 6 points × 1.0 = 6 points, and a light load is 4 points × 0.8 = 3.2 points, more accurately reflecting the impact of load on passing priority. The task urgency factor includes urgent, urgent, routine, and urgent time-efficiency corrections. For example, an urgent vehicle requiring completion within 30 minutes has a time-efficiency coefficient of 1.5, while a vehicle requiring completion within 1 hour has a coefficient of 1.2. The additional score = urgency level bonus score × time-efficiency coefficient. For example, an urgent vehicle has a bonus score of 10 points × 1.5 = 15 points, an urgent vehicle has 6 points × 1.2 = 7.2 points, and a routine vehicle has 2 points × 1.0 = 2 points, highlighting the impact of task timeliness on priority. The global congestion contribution factor uses a digital twin engine to obtain the overall tunnel congestion status, such as the length of congested sections and the number of congested vehicles, to calculate the contribution of each vehicle to congestion. If a vehicle's speed is below 60% of the tunnel speed limit and it is in the core area of ​​a congested section, the contribution coefficient is 1.3; if the speed is normal and it is in a clear section, the coefficient is 0.7. The congestion correction score = (10 - vehicle congestion contribution score) × contribution coefficient. The higher the congestion contribution score, the lower the correction score, preventing high-contribution vehicles from occupying too much traffic resources. The final priority score is calculated as follows: vehicle load base score × 0.3 + task urgency bonus score × 0.5 + congestion correction score × 0.2. The weights can be dynamically adjusted based on the scenario. The total score is divided into four levels to ensure that the priority calculation better reflects the overall operational needs of the tunnel. It's worth noting that vehicle load affects vehicle speed, braking distance, and thus traffic flow smoothness; the direction of travel determines the flow of vehicles within the tunnel, and the interaction of vehicles in different directions affects traffic efficiency; task urgency reflects the importance of the task carried by the vehicle, with vehicles carrying urgent tasks requiring priority passage; distance to the passing point relates to the convenience and safety of passing; and the contribution to the overall tunnel congestion status considers the vehicle's priority from the perspective of the overall traffic situation. Combining these calculation factors allows for a more scientific and accurate reflection of vehicle priority in complex traffic environments.

[0043] The dimensional data acquisition phase is divided into several sub-phases. The fluctuation trends of dimensional data in each sub-phase and the entire acquisition phase are collected. If the corresponding fluctuation trends are inconsistent, the sub-phase is marked as a floating phase; if the corresponding fluctuation trends are consistent, the sub-phase is marked as a stable phase. During a floating phase, the interval between the time of dimensional data trend analysis and the end of the floating phase is collected. During a stable phase, the numerical increase span of corresponding dimensional data from multiple stable phases is collected during dimensional data trend analysis. If the time interval in a floating phase exceeds the interval threshold, or the numerical increase span in a stable phase exceeds the span threshold, it is inferred that the lag analysis during dimensional data acquisition is abnormal. The service demand assessment platform shortens the dimensional data acquisition prediction cycle, i.e., real-time fluctuation prediction is performed during the floating phase, and a critical value for the numerical fluctuation span is set during the stable phase; if exceeded, demand prediction is performed promptly. If the time interval in a floating phase does not exceed the interval threshold, and the numerical increase span in a stable phase does not exceed the span threshold, it is inferred that the lag analysis during dimensional data acquisition is normal. The vehicle-passing algorithm module acquires real-time operational status data of the current vehicle and vehicles within a 500-meter radius using a fusion positioning system. This data includes vehicle location, speed, direction, type, and load information. Simultaneously, it receives global tunnel status information from the digital twin engine module, including congestion levels in different tunnel sections, traffic flow at key cross-sections, work area distribution, and emergency event information. Both types of data are standardized to construct a real-time environmental state vector of dimension n×m, where n represents the number of vehicles and m is the sum of the individual vehicle's state dimension and the global state dimension, providing a data foundation for subsequent decision-making. The reinforcement learning model dynamically calculates the perception distance threshold based on the real-time environmental state vector. The threshold is set to 50 meters when the surrounding vehicle density is high and 100 meters when the density is low. When the actual distance between the current vehicle and surrounding vehicles is less than the perception distance threshold, a risk of vehicle-passing conflict is identified, triggering a distributed collaborative decision-making process. If the actual distance is greater than the threshold, the decision-making process is not triggered, and the vehicle continues on its original route.

[0044] It is worth noting that the vehicle, acting as an autonomous intelligent agent, inputs the constructed environmental state vector into a locally deployed reinforcement learning policy network. Based on a multi-objective optimization function, the network outputs preliminary traffic control commands, such as "accelerate to 25 km / h," "decelerate to 10 km / h," and "stop and wait." Simultaneously, a dynamic priority allocation unit, based on predefined weight rules (load 0.3, urgency 0.5, congestion contribution 0.2), calculates the real-time priority scores of the current vehicle and surrounding conflicting vehicles, generating a priority ranking table.

[0045] During this process, the current vehicle broadcasts its initial traffic instructions and its real-time priority score to surrounding conflicting vehicles via vehicle-to-vehicle (V2V) communication technology, while simultaneously receiving instructions and scores broadcast by other vehicles. If multiple vehicles have conflicting initial instructions, such as vehicle A's instruction being "acceleration priority" and vehicle B's instruction also being "acceleration priority," posing a collision risk, arbitration is conducted based on the real-time priority scores: the vehicle with the highest priority score is determined as the priority vehicle, and the remaining vehicles are determined as yielding vehicles. If the priority scores are the same, the yielding vehicle is determined by vehicle type or direction of travel, with larger vehicles taking precedence over smaller vehicles, and mainline vehicles taking precedence over sideline vehicles. The yielding vehicle executes the corresponding yielding action based on the arbitration result, such as deceleration, stopping, or lane changing, and provides real-time feedback on its execution progress via V2V. The priority vehicle smoothly passes through the conflict area according to the instructions output by the reinforcement learning strategy network. During execution, the fusion positioning system monitors vehicle position changes in real time. If it detects that the yielding action is not executed properly or the yielding vehicle's deceleration is insufficient, it immediately feeds back to the module to readjust the action instructions. For vehicles temporarily stopped due to a collision, their locally deployed reinforcement learning policy network continuously receives real-time environmental state vectors to predict whether starting the collision will trigger a new conflict, such as when the distance to oncoming vehicles is less than a safe threshold. If no conflict is predicted, or if a system release command is received from the digital twin engine module, the vehicle initiates the collision, slowly accelerating and merging into the normal traffic flow. If a conflict is predicted, the vehicle remains stopped until the collision conditions are met. After the collision is completed, the module transmits the results, including collision time, safe distance compliance, energy consumption data, and conflict resolution effectiveness, back to the digital twin engine module, edge computing cluster, and data collaborative governance module in real time. Based on this feedback data, the reinforcement learning training unit incrementally trains the policy network, continuously optimizing the multi-objective optimization function parameters to improve the accuracy and efficiency of subsequent decisions.

[0046] In the execution process, the integrated positioning system provides high-precision vehicle status data for the data acquisition stage, ensuring the accuracy of the environmental status vector; the digital twin engine provides support for global status information, affecting the calculation of perception distance thresholds and the judgment of vehicle start-up; the edge computing cluster realizes the localized deployment of reinforcement learning strategy network, ensuring low latency in the generation of action commands; the data collaborative governance module verifies vehicle V2V communication permissions through the smart contract sub-module and records execution logs through the blockchain sub-module, forming a cross-module collaborative closed loop of data support, decision execution, and feedback optimization.

[0047] The edge computing cluster includes: A lightweight AI inference engine, deployed on edge nodes at the tunnel site, supporting TensorFlow Lite or ONNXRuntime frameworks; The data encryption unit encrypts sensitive data using the national cryptographic SM4 algorithm before uploading it to the cloud.

[0048] Specifically, the lightweight AI inference engine is deployed at edge nodes in the tunnel, such as tunnel entrance and exit server rooms and temporary computing sites in the middle section. Addressing the computing power requirements and hardware resource limitations of tunnel scenarios, it adopts a lightweight design, supports mainstream AI frameworks, and provides low-latency, high-performance AI inference services for various system modules. The engine natively supports two major lightweight AI frameworks: TensorFlow Lite (Google's lightweight deep learning framework) and ONNX Runtime (Open Neural Network Exchange Format Runtime). It can directly load machine learning models (such as personnel density prediction models and equipment failure risk models) and reinforcement learning policy network models from the vehicle passing algorithm module, without requiring complex model format conversions, reducing cross-framework adaptation costs. Furthermore, considering the limited hardware resources of edge nodes, the engine has automatic lightweight model optimization capabilities: through model pruning, quantization, and knowledge distillation techniques, it compresses the model size and improves inference speed while ensuring inference accuracy, adapting to the low computing power and low power consumption requirements of edge nodes. For example, a 200MB original reinforcement learning model for passing vehicles can be compressed to less than 50MB after optimization, and the inference latency can be reduced from 50 milliseconds to less than 10 milliseconds. The engine supports parallel inference for multi-module AI tasks, and can simultaneously process location data parsing tasks from the fusion positioning system, risk prediction tasks from the digital twin engine, and action decision tasks from the passing vehicle algorithm module. To avoid increased latency caused by multiple tasks competing for computing resources, the engine has a built-in intelligent resource scheduling algorithm that dynamically allocates computing resources based on task priority (e.g., passing vehicle action decision task has the highest priority, followed by location data parsing, and then risk prediction) and the real-time computing load of edge nodes. For example, when the passing vehicle algorithm module triggers a distributed decision-making process, the engine automatically allocates 70% of the GPU computing power to the passing vehicle action decision inference task to ensure low latency in instruction generation; when there is no passing vehicle decision task, the surplus computing power is allocated to the risk prediction task to improve prediction efficiency. In addition, the engine has a task caching and preloading mechanism. For high-frequency repetitive inference tasks, intermediate results during the inference process are cached to reduce redundant calculations; for tasks that are about to be triggered, the corresponding models and parameters are preloaded to shorten the task startup latency. After completing AI inference, the engine transmits the inference results back to the corresponding modules in real time via a high-speed data interface: the location data parsing results are transmitted to the fused positioning system and the digital twin engine for correcting dynamic entity mapping; the risk prediction results are transmitted to the digital twin engine for generating optimized scheduling strategies; and the inference results of the vehicle meeting action commands are transmitted to the vehicle meeting algorithm module for decision execution. Simultaneously, the engine monitors its own operational status in real time and uploads the status data to the digital twin engine's monitoring panel. When it detects that the inference latency exceeds a preset threshold, the hardware temperature is too high, or the model loading fails, it immediately triggers an alert, notifying maintenance personnel to troubleshoot the fault and ensure stable engine operation.

[0049] The data encryption unit employs national cryptographic algorithms to encrypt sensitive data generated during tunnel scheduling, such as personnel location privacy data, core equipment operating parameters, and scheduling instructions. This ensures the security and confidentiality of data uploaded from edge nodes to the cloud. Through preset sensitive data identification rules, it automatically identifies sensitive information in various types of data generated by edge nodes. High-sensitivity data includes personnel identification and location trajectory association data, core equipment operating parameters, and emergency scheduling instructions; medium-sensitivity data includes tunnel traffic flow data and material transfer records; and low-sensitivity data includes tunnel environmental monitoring data and edge node operating status data. Differentiated encryption strategies are used for data of different sensitivity levels: high-sensitivity data uses both the national cryptographic SM4 algorithm and digital signatures for dual protection; medium-sensitivity data uses only the SM4 algorithm; and low-sensitivity data uses simple verification, balancing security with encryption efficiency. The national cryptographic SM4 algorithm (block cipher algorithm) possesses high-strength encryption capabilities, resisting common cryptographic attacks such as brute-force and differential attacks, and meets the requirements of Level 3 or higher of the National Information Security Protection System, making it suitable for the encryption needs of sensitive data in tunnel scheduling. The unit employs an edge node-cloud two-way key management mechanism to ensure the security of encryption keys: The cloud key management center (KMC) periodically generates SM4 algorithm encryption keys, which are then encrypted using a hardware encryption module (HSM) and distributed to the data encryption units of each edge node. Upon receiving the keys, the edge nodes store them in a secure encryption chip to prevent unauthorized access. When an edge node is replaced, a key's usage period expires, or there is a risk of key leakage, the cloud KMC triggers a key update process, issuing a new key to the edge node and simultaneously instructing the old key to be immediately destroyed. When destroying old keys, edge nodes use multiple overwrite and physical deletion methods to ensure the keys cannot be recovered. After sensitive data is generated, the data encryption unit retrieves the corresponding key from the secure encryption chip and uses the SM4 algorithm to encrypt the data in groups, generating encrypted ciphertext. For highly sensitive data, an additional digital signature is performed using the edge node's digital certificate to ensure data integrity and the authenticity of the sender's identity. After encryption, the ciphertext and digital signature are packaged and uploaded to the cloud via a 5G encrypted transmission channel. Upon receiving the data, the cloud retrieves the corresponding key to decrypt it, verifies the digital signature, and stores or uses the data only after confirming its accuracy.

[0050] Example 2: See Figure 2 A tunnel scheduling and command method based on digital twins and distributed decision-making includes: Construct a three-dimensional digital twin model of the tunnel space, integrate BIM, GIS, multi-source environmental sensor data and dynamic entity mapping, and integrate machine learning models to predict personnel gathering trends and equipment failure risks in real time, and dynamically generate optimized scheduling strategies. Blockchain distributed ledger records scheduling decisions and material flow data to ensure that the data is tamper-proof and auditable throughout the entire process; The integrated positioning system is used to achieve positioning and low-latency communication in the tunnel. The signal adaptive switching mechanism is used to automatically switch to UWB standalone mode in the 5G signal coverage blind spot. The adaptive learning vehicle meeting algorithm is implemented, which enables the vehicle agent to optimize the traffic strategy in real time through a reinforcement learning model, and adjusts the avoidance threshold and priority based on historical traffic data and real-time environmental perception. It leverages edge computing clusters to process real-time data locally, accelerates the execution of machine learning models through an AI inference engine, and collaborates with a blockchain module to ensure data security.

[0051] Specifically, the process begins with collecting core data throughout the tunnel's entire lifecycle, including BIM data from the construction phase and GIS data from the geospatial dimension. During operation, multi-source environmental sensors deployed within the tunnel collect dynamic environmental data in real time. Simultaneously, dynamic status data of personnel, equipment, and vehicles within the tunnel are integrated to construct a 3D digital twin model, achieving a comprehensive digital mapping of all elements of the physical tunnel. Multiple machine learning models are integrated into the 3D digital twin model, including models predicting personnel gathering trends and equipment failure risks. Through real-time model computation, the safety status of personnel and equipment within the tunnel is continuously monitored, and potential risks are promptly identified. Based on model predictions and the tunnel's real-time operational needs, optimized scheduling strategies for multiple scenarios are automatically generated. For example, when the predicted risk of personnel gathering in a construction area exceeds the standard, a scheduling instruction to divert personnel to a backup work area is generated; when the predicted probability of equipment failure exceeds 90%, a scheduling plan to prioritize maintenance teams to handle the faulty equipment is generated. Furthermore, the strategy dynamically iterates based on real-time feedback from the physical tunnel, ensuring that the scheduling plan always adapts to the actual operational status of the tunnel. The core data types that need to be recorded are clearly defined, including scheduling decision data and material flow data, ensuring that data covers the entire lifecycle of scheduling instructions from generation to execution. Utilizing blockchain distributed ledger technology, key collected data is packaged into blocks in timestamp order and stored synchronously across multiple nodes, preventing data loss due to single-node failure. Simultaneously, the data is encrypted, with each node accessing it only through its dedicated private key, ensuring unauthorized data tampering. For example, once a material allocation record is written to the ledger, any attempt by a node to modify the material receipt time requires consensus verification from over 51% of the nodes, significantly reducing the risk of data tampering. A data auditing mechanism is established, supporting rapid querying of complete data flow records using data keywords, including data generation and storage nodes, forming a visualized traceability chain. Furthermore, the system automatically generates data audit reports, statistically analyzing key indicators such as dispatch instruction execution efficiency and material turnover rate, meeting the compliance audit requirements for tunnel operations and providing data support for dispatch process optimization. Furthermore, UWB (Ultra-Wideband) positioning base stations are deployed within the tunnel, planned according to density requirements for straight sections and curved sections to ensure full positioning signal coverage. Personnel are equipped with UWB positioning tags, and equipment and vehicles are equipped with UWB positioning terminals. The coordinates of personnel, equipment, and vehicles are calculated by receiving signals transmitted by the tags / terminals through the UWB base stations. Simultaneously, 5G network slicing technology is used to establish a dedicated communication channel for positioning data transmission, ensuring low latency and meeting the real-time requirements of dispatching. A tunnel map is pre-imported into the system, marking areas with weak 5G signal coverage or blind spots (such as the middle section of long tunnels or deeply buried tunnel areas), and setting 5G signal strength thresholds.Signal monitoring sensors deployed within the tunnel collect 5G signal strength in real time. When the 5G signal is detected to be below a threshold for 10 consecutive seconds, or when vehicles or personnel enter a pre-defined 5G blind zone, the system automatically triggers signal switching, ceasing 5G transmission of positioning data and switching to UWB independent mode. This mode directly collects and transmits positioning data via UWB base stations, ensuring uninterrupted positioning. When the 5G signal recovers, the system automatically switches back to 5G mode, balancing positioning accuracy and communication efficiency. Simultaneously, a reinforcement learning model is constructed with multiple optimization objectives, including traffic efficiency, safe distance, and energy consumption. Historical tunnel traffic meeting data is collected to train the model: the real-time environment during vehicle encounters is defined as the state space, and feasible vehicle operations are defined as the action space. By calculating the reward values ​​corresponding to different actions, the model parameters are iteratively optimized, enabling the vehicle to possess the attributes of an autonomous decision-making agent. During operation, the vehicle's intelligent agent collects real-time positioning data of surrounding vehicles and tunnel traffic flow status, inputting this data into the reinforcement learning model to dynamically output the optimal traffic strategy. Meanwhile, based on historical traffic data and real-time environmental perception, the system automatically adjusts the avoidance threshold: for example, when the road surface is slippery due to rain, the safe passing distance threshold is increased from 5 meters to 8 meters; when the tunnel is congested, the distance threshold that triggers the passing decision is shortened from 100 meters to 50 meters to ensure passing safety. Based on the real-time vehicle status, a passage priority score is calculated, with factors including vehicle load, task urgency, and contribution to tunnel congestion. When multiple vehicles have conflicting passing instructions, the system arbitrates based on the priority score to determine the avoidance party and the priority party. The avoidance party performs deceleration or stopping, while the priority party passes through the conflict area as instructed, avoiding the risk of a collision. Edge computing clusters are deployed at tunnel entrances / exits or in the middle section to transmit raw UWB positioning data collected by the fusion positioning system, multi-source environmental sensor data, and vehicle status data required by the passing algorithm to the edge clusters in real time. The cluster utilizes a built-in lightweight AI inference engine to locally execute data processing tasks: performing TDOA algorithm inference on raw UWB data to output positioning results; accelerating inference on personnel gathering and equipment failure prediction models issued by the digital twin engine to output risk warning results; and inferring the reinforcement learning model of the vehicle passing algorithm to generate vehicle passage instructions, controlling data processing latency and avoiding latency losses caused by data uploading to the cloud. The AI ​​inference engine uses model lightweighting technology to compress the model size and improve inference speed without sacrificing inference accuracy. Simultaneously, the engine dynamically allocates computing resources based on task priority. For example, in vehicle passing scenarios, 70% of GPU computing power is allocated to vehicle passing model inference to ensure efficient instruction generation; when there are no urgent tasks, surplus computing power is allocated to risk prediction to optimize overall scheduling efficiency.The edge computing cluster encrypts the processed sensitive data using the national cryptographic algorithm SM4, and then uploads the encrypted data to the blockchain distributed ledger. At the same time, the data processing logs are synchronized to the ledger, enabling secure traceability of the entire data process from processing to storage, ensuring that sensitive data is not leaked or tampered with, and meeting the data security requirements of tunnel scheduling.

[0052] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and all fall within the scope of the technical solution.

Claims

1. A tunnel scheduling and command system based on digital twins and distributed decision-making, characterized in that, include: The digital twin engine module is used to construct a three-dimensional digital twin model of the tunnel space. The three-dimensional digital twin model integrates BIM, GIS, multi-source environmental sensor data and dynamic entity mapping, outputs personnel gathering trends, equipment failure risks and traffic flow dynamics, and dynamically generates optimized scheduling strategies based on the prediction results. The data collaborative governance module is used to receive the predicted data output by the digital twin engine module, perform correlation analysis with the historical scheduling decision data recorded in the distributed ledger, convert the analysis results into material scheduling instructions through smart contracts, and feed them back to the digital twin engine module to update the scheduling strategy. The integrated positioning system is used to upload UWB positioning data of personnel and equipment in the tunnel to the edge computing cluster in real time through 5G network slicing, receive the positioning parsing results processed by the edge computing cluster, and synchronize them to the digital twin engine module to correct the dynamic entity mapping. The vehicle meeting algorithm module is used to obtain real-time traffic flow data and predicted avoidance thresholds from the digital twin engine module, combine them with the adjacent vehicle positioning information provided by the fusion positioning system, generate a decentralized traffic strategy, and send the strategy execution result to the transportation equipment after authorization verification through the smart contract of the data collaborative governance module. Edge computing clusters are used for localized processing of raw positioning data from fusion positioning systems. They output positioning analysis results through an AI inference engine, accelerate inference of machine learning models issued by the digital twin engine module, and send the results back to the vehicle meeting algorithm module to optimize decision-making timeliness.

2. The tunnel scheduling and command system based on digital twins and distributed decision-making according to claim 1, characterized in that, The digital twin engine module includes: A multi-source data fusion unit is used to synchronously update environmental sensor data, equipment status data, and personnel positioning data within the tunnel to the three-dimensional digital twin model; The prediction model training unit trains a machine learning model based on historical data. The output of the machine learning model includes a population density heatmap, equipment failure probability, and traffic congestion warning.

3. The tunnel scheduling and command system based on digital twins and distributed decision-making according to claim 1, characterized in that, The data collaborative governance module includes: The blockchain sub-module uses distributed ledger technology to store scheduling instruction execution logs, material flow records, and permission change information; The smart contract submodule has a preset scheduling rule library, which automatically performs resource allocation, permission verification, or task issuance when preset conditions are triggered.

4. The tunnel scheduling and command system based on digital twins and distributed decision-making according to claim 1, characterized in that, The fusion positioning system includes: The 5G network slicing unit is configured to prioritize low latency in the transmission of scheduling instructions; UWB positioning units enable the location of personnel and equipment through UWB base stations deployed inside tunnels; Among them, based on the predefined signal coverage area of ​​the tunnel map, the signal strength is monitored through real-time sensor data, and the 5G signal is automatically switched to UWB positioning mode when it is lower than the preset threshold.

5. The tunnel scheduling and command system based on digital twin and distributed decision-making according to claim 1, characterized in that, The vehicle meeting algorithm module includes: The reinforcement learning training unit uses vehicle traffic efficiency, safe distance and energy consumption as optimization goals to iteratively update the avoidance strategy; The dynamic priority allocation unit adjusts the passage priority in real time based on vehicle load, direction of travel, urgency of the task, tunnel congestion, and distance to the passing point.

6. The tunnel scheduling and command system based on digital twin and distributed decision-making according to claim 1, characterized in that, The edge computing cluster includes: A lightweight AI inference engine, deployed on edge nodes at the tunnel site, supporting TensorFlow Lite or ONNX Runtime framework; The data encryption unit encrypts sensitive data using the national cryptographic SM4 algorithm before uploading it to the cloud.

7. The tunnel scheduling and command system based on digital twin and distributed decision-making according to claim 5, characterized in that, The execution of the meeting algorithm module includes: Real-time acquisition of the operating status data of the current vehicle and surrounding vehicles, and receipt of global tunnel status information, together constructing a real-time environmental status vector; Based on the perception distance threshold dynamically calculated by the machine learning model, it is determined whether to trigger the distributed collaborative decision-making process; The current vehicle, acting as an autonomous intelligent agent, inputs the environmental state vector into a locally deployed reinforcement learning policy network, outputs preliminary passage instructions, and calculates the real-time priority scores of the current vehicle and conflicting vehicles based on predefined weight rules. The vehicle broadcasts instructions and scores via vehicle-to-vehicle (V2V) communication, and in the event of instruction conflicts, arbitrates the vehicle to give way based on the real-time priority score. The vehicle that is giving way performs an evasive maneuver, and the vehicle with priority passes through the conflict area according to instructions. For parked vehicles, the startup operation is performed when the policy network predicts that startup will not cause a conflict, or when a system release instruction is received.

8. The tunnel scheduling and command system based on digital twins and distributed decision-making according to claim 7, characterized in that, The reinforcement learning policy network is trained iteratively using vehicle traffic efficiency, safe distance and energy consumption as multi-objective optimization functions. Furthermore, the calculation factors for the real-time priority score include at least: vehicle load, driving direction, task urgency, distance from the passing point, and their contribution to the overall congestion status of the tunnel.

9. The tunnel scheduling and command system based on digital twin and distributed decision-making as described in claim 8, characterized in that, The traffic efficiency includes the average travel time of vehicles from the tunnel entrance to the tunnel face, the number of vehicles passing through the key sections of the tunnel per unit time, the average waiting time of vehicles at the conflict point, and the utilization rate of the avoidance facilities. The safe distance includes the vehicle dynamic safe distance, the safe margin buffer zone, and the human-vehicle interaction safe distance; The energy consumption includes driving energy consumption, total fuel consumption, and energy costs.

10. A tunnel scheduling and command method based on digital twins and distributed decision-making, characterized in that, include: Construct a three-dimensional digital twin model of the tunnel space, integrate BIM, GIS, multi-source environmental sensor data and dynamic entity mapping, and integrate machine learning models to predict personnel gathering trends and equipment failure risks in real time, and dynamically generate optimized scheduling strategies. Blockchain distributed ledger records scheduling decisions and material flow data to ensure that the data is tamper-proof and auditable throughout the entire process; The integrated positioning system is used to achieve positioning and low-latency communication in the tunnel. The signal adaptive switching mechanism is used to automatically switch to UWB standalone mode in the 5G signal coverage blind spot. The adaptive learning vehicle meeting algorithm is implemented, which enables the vehicle agent to optimize the traffic strategy in real time through a reinforcement learning model, and adjusts the avoidance threshold and priority based on historical traffic data and real-time environmental perception. It leverages edge computing clusters to process real-time data locally, accelerates the execution of machine learning models through an AI inference engine, and collaborates with a blockchain module to ensure data security.

Images