Tunnel management method and system based on digital twinning

Abstract

The invention relates to the technical field of digital twinning, and particularly discloses a tunnel management method and system based on digital twinning. The system comprises an equipment internet-of-things sensing layer, a data fusion processing layer, a digital twin model layer, an intelligent analysis decision-making layer and a cooperative control execution layer, and realizes global monitoring, intelligent analysis and cooperative control of a tunnel state by constructing high-fidelity real-time mapping of a physical tunnel and a virtual model. Therefore, the reliability, the safety and the operation and maintenance efficiency of tunnel management are improved.

Description

Technical Field

[0001] This invention belongs to the field of digital twin technology, specifically relating to a tunnel management method and system based on digital twins. Background Technology

[0002] As a core component of modern transportation infrastructure, tunnel engineering plays a crucial role in ensuring smooth road networks and public safety through its safe and efficient operation. With the rapid development of technologies such as the Internet of Things and big data, tunnel operation and management is gradually evolving towards digitalization and intelligence, with the construction of high-fidelity, real-time interactive digital twin systems becoming a significant industry trend. During tunnel operation, the status monitoring, coordinated control, and emergency response capabilities of various electromechanical equipment form the foundation of system reliability.

[0003] Among them, the tunnel management method based on digital twins aims to achieve integrated monitoring and intelligent decision-making of tunnel structure, equipment, and environment by constructing a real-time mapping relationship between the physical tunnel and the virtual model. This method relies on data acquisition, communication transmission, and integrated analysis of various terminal devices within the tunnel to support key functions such as equipment status assessment, risk warning, and scheduling optimization.

[0004] In existing technologies, the numerous electromechanical terminal devices deployed within tunnels generally suffer from low levels of intelligence and inconsistent communication protocols, creating information silos and making it difficult to perceive and manage equipment status in real time. In emergency scenarios, the lack of effective equipment linkage mechanisms and intelligent decision support leads to slow emergency response, which can easily trigger secondary accidents and increase operational safety risks.

[0005] Meanwhile, traditional operation and maintenance models rely on manual inspections and experience-based judgment, making it difficult to guarantee the timeliness of equipment defect identification and the comprehensiveness of potential hazard investigation, further leading to high operation and maintenance costs and low operational efficiency. Therefore, there is an urgent need to build a tunnel management method that can achieve equipment data fusion, intelligent analysis, and collaborative control to improve the overall reliability, safety, and operation and maintenance efficiency of tunnel systems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a tunnel management method and system based on digital twins, so as to overcome the problems in the prior art caused by the low level of intelligence of tunnel electromechanical terminal equipment and the lack of unified communication protocols, resulting in information silos, and the resulting difficulties in real-time perception and unified management of equipment status, slow emergency response, and low efficiency due to reliance on manual experience in operation and maintenance.

[0007] This invention provides a tunnel management system based on digital twins. This system establishes a high-fidelity, real-time, bidirectional mapping relationship between the physical tunnel and its virtual model, enabling integrated monitoring, intelligent analysis, and collaborative control of the entire tunnel's status. The system includes an equipment IoT sensing layer, a data fusion processing layer, a digital twin model layer, an intelligent analysis and decision-making layer, and a collaborative control execution layer.

[0008] The device IoT sensing layer is responsible for collecting operational status data and environmental parameters of various electromechanical terminal equipment in the tunnel;

[0009] The data fusion processing layer parses, cleans, and standardizes heterogeneous data from different protocol interfaces, and completes the spatiotemporal alignment and feature extraction of multi-source data; the digital twin model layer dynamically drives the virtual tunnel model based on the processed data, realizing the real-time mirroring of physical entities in virtual space;

[0010] The intelligent analysis and decision-making layer uses the panoramic status information output by the digital twin model to conduct equipment health assessment, risk prediction, and emergency response strategy generation.

[0011] The collaborative control execution layer then implements precise linkage control of relevant equipment groups within the tunnel based on decision-making instructions.

[0012] Furthermore, the equipment IoT sensing layer includes various sensor nodes and smart terminals deployed within the tunnel. These nodes and terminals are functionally divided into structural monitoring units, environmental monitoring units, traffic flow monitoring units, and equipment status monitoring units.

[0013] The structural monitoring unit uses a fiber optic grating sensor array and distributed strain gauges to continuously monitor the stress, strain and displacement changes of the tunnel lining at a sampling frequency of 1 time / s.

[0014] The environmental monitoring unit integrates temperature and humidity sensors, visibility detectors, and multi-gas concentration analyzers to collect real-time data on temperature, humidity, visibility, and carbon monoxide and nitrogen oxide concentrations within the tunnel.

[0015] The traffic flow monitoring unit uses millimeter-wave radar and high-definition video cameras to acquire information on traffic flow, average vehicle speed, and vehicle density in the tunnel every 5 seconds.

[0016] The equipment status monitoring unit uses built-in or external intelligent acquisition modules to acquire in real time the voltage, current, power factor, running time, and fault code status of key electromechanical equipment such as fans, lighting fixtures, fire pumps, and variable message signs.

[0017] Furthermore, the data fusion processing layer includes a protocol adaptation module, a data cleaning module, and a feature fusion engine. The protocol adaptation module has a built-in parsing library for at least 8 industrial fieldbus protocols and 3 wireless IoT communication protocols, which can automatically identify the communication protocols of access devices and convert their data into the system's internal standard data format.

[0018] The data cleaning module detects and removes outliers from the raw data stream and uses a mean shift algorithm based on a sliding time window to imput and repair missing data, ensuring the continuity and integrity of the data stream.

[0019] The feature fusion engine performs spatiotemporal alignment processing on the cleaned multi-source heterogeneous data. It uses the high-precision BeiDou satellite positioning system and network time protocol to give each piece of data a unified timestamp and spatial coordinates, and uses a graph neural network-based association mining algorithm to extract deep coupling features between equipment status parameters and environmental factors and traffic load.

[0020] Furthermore, the digital twin model layer consists of a geometric model, a physical model, and a behavioral rule model. The geometric model is constructed using lightweight Building Information Modeling (BIM) technology to accurately reproduce the tunnel's internal structure, equipment layout, and pipeline routing.

[0021] The physical model incorporates the mechanical constitutive relationship of the tunnel lining, the aerodynamic equations, and the aging and wear patterns of the equipment. It is used to simulate the mechanical response of the tunnel structure under vehicle loads and environmental effects, as well as the degradation trend of equipment performance over time.

[0022] The behavioral rule model defines the interaction logic and control strategy between various entities in the tunnel system under the triggering of specific events, such as the linkage rules between fire alarm signals and fan start-up, lighting enhancement, and traffic guidance screen information release.

[0023] This model layer ensures dynamic consistency between the virtual model and the physical tunnel state through a real-time data-driven engine with microsecond-level time synchronization accuracy.

[0024] Furthermore, the intelligent analysis and decision-making layer integrates an equipment health assessment module, a multi-step risk prediction module, and an emergency response strategy generation module. The equipment health assessment module, based on real-time data and historical operating records uploaded by the equipment status monitoring unit, uses a deep residual network to construct an equipment health index calculation model. This model can identify early signs of equipment performance degradation and predict its remaining service life, automatically issuing an early warning when the health index falls below a preset threshold of 0.85.

[0025] The multi-step risk prediction module uses a long short-term memory network to perform time series analysis on the fused tunnel environment, traffic flow and structural monitoring data to predict the probability of congestion, fire or structural anomaly in the tunnel within the next 30 minutes. A high-level warning is activated when the risk probability is greater than 70%.

[0026] The emergency response strategy generation module is based on reinforcement learning algorithms. It performs parallel simulations of various potential emergency response plans in a digital twin model, evaluates the comprehensive utility of each plan in terms of execution effect, resource consumption and secondary risks, and finally selects the strategy with the highest utility value as the recommended plan output.

[0027] Furthermore, the collaborative control execution layer receives control commands from the intelligent analysis and decision-making layer and distributes them to various execution terminals at the tunnel site through a distributed control network. This layer has command verification and safety interlocking functions to ensure that control commands comply with equipment operating procedures and system safety policies.

[0028] For example, upon receiving an instruction to start the fans for longitudinal ventilation, the floor will simultaneously check the status of the fire detectors and traffic flow in the tunnel. Only when it is confirmed that there are no vehicles blocking the ventilation duct and the fire alarm signal is real and valid will the start command be sent to the designated fan group.

[0029] Meanwhile, this layer supports dynamic adjustment of control strategies, and can fine-tune control parameters to optimize overall performance based on the real-time control effect feedback from the digital twin model.

[0030] Furthermore, as one embodiment of the present invention, the training process of the deep residual network in the equipment health assessment module is as follows.

[0031] First, a large amount of historical operating data covering normal equipment conditions, various typical failure modes, and performance degradation processes is collected to form a training sample set.

[0032] Then, the sample data is normalized, and multi-dimensional feature vectors including time-domain features, frequency-domain features, and time-frequency-domain features are extracted.

[0033] Next, a deep residual network model with 15 hidden layers is constructed. The number of nodes in the input layer is equal to the dimension of the feature vector, and the output layer has 1 node, which corresponds to the health index of the device. The value of this index ranges from 0 to 1.

[0034] Finally, the backpropagation algorithm and the adaptive moment estimation algorithm were used to train the network model until the root mean square error of the model on the independent validation set was less than 0.05.

[0035] Furthermore, as another embodiment of the present invention, the construction and operation process of the long short-term memory network in the multi-step risk prediction module is as follows. The network includes two stacked long short-term memory layers and one fully connected output layer.

[0036] The input data consists of tunnel environment data sequences, traffic flow data sequences, and key equipment status data sequences within the past 60-minute time window, preprocessed by the data fusion processing layer. The network learns long-term dependencies in the historical sequences and outputs a probability sequence of specific risk events occurring within the tunnel in the next 6 time steps, i.e., the next 30 minutes.

[0037] The loss function used to train the network is weighted cross-entropy, which assigns higher weights to high-risk samples to improve the model's sensitivity in predicting minority dangerous events.

[0038] Furthermore, as another embodiment of the present invention, the specific implementation framework of the reinforcement learning algorithm in the emergency response strategy generation module is as follows.

[0039] The tunnel emergency response environment is modeled as a Markov decision process, where the state space consists of real-time tunnel panoramic state variables output by the digital twin model, the action space corresponds to all executable control command combinations, and the reward function comprehensively considers the accident control effect, rescue time, impact on normal traffic, and resource consumption cost.

[0040] The Markov decision process is solved using a proximal policy optimization algorithm. The agent learns an optimal policy network capable of generating efficient handling strategies in various complex accident scenarios through extensive interactive simulation training with a digital twin environment.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. This invention, by constructing a system architecture integrating equipment IoT sensing, data fusion processing, digital twin modeling, intelligent analysis and decision-making, and collaborative control execution, completely breaks down the information silos caused by inconsistent protocols among various electromechanical terminal devices within the tunnel, achieving holographic perception and unified management of the entire tunnel's status. The data fusion processing layer provides a high-quality, highly consistent data foundation for upper-layer applications through in-depth processing and feature fusion of multi-source heterogeneous data.

[0043] 2. This invention utilizes a digital twin model layer to achieve high-fidelity, real-time, bidirectional mapping between the physical tunnel and the virtual space, enabling managers to intuitively and comprehensively grasp the tunnel's operational status. The embedded physical model and behavioral rule model allow the system to not only reflect the current state but also perform high-precision simulation and prediction of structural response, equipment performance degradation, and the evolution of emergencies, providing strong support for forward-looking operation and maintenance and proactive safety control.

[0044] 3. The intelligent analysis and decision-making layer of this invention integrates deep learning-based equipment health assessment, multi-step risk prediction, and reinforcement learning-based emergency response strategy generation, thereby improving the intelligence level of tunnel management. The system can identify potential equipment failures in advance, predict operational risks, and quickly generate and simulate the optimal handling plan in the event of an emergency, greatly shortening the emergency response time and avoiding secondary accidents that may be caused by slow decision-making.

[0045] 4. The collaborative control execution layer of this invention ensures that intelligent decisions can be implemented accurately and safely. Its instruction verification and safety interlocking mechanisms guarantee the reliability of control actions and the overall security of the system, while the dynamic optimization capability based on digital twin feedback further improves the accuracy of control and the energy efficiency of system operation. Overall, this invention enhances the reliability, safety, and operation and maintenance efficiency of tunnel systems, and reduces the total life cycle operating costs. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall technical solution architecture of the tunnel management system based on digital twin proposed in this invention;

[0047] Figure 2 This is a schematic diagram of the core principle framework of data fusion processing and intelligent analysis and decision-making in this invention;

[0048] Figure 3 This is a schematic diagram of the core principle framework of the digital twin model layer in this invention;

[0049] Figure 4 This is a flowchart of the main stages of the intelligent analysis and decision-making layer in this invention; Detailed Implementation

[0050] Please refer to the attached document. Figure 1 This embodiment details the implementation of a tunnel management system based on digital twins. This system establishes a high-fidelity, real-time, two-way mapping relationship between the physical tunnel and its virtual model, enabling integrated monitoring, intelligent analysis, and collaborative control of the entire tunnel's status.

[0051] The system consists of five core layers: device IoT sensing layer, data fusion and processing layer, digital twin model layer, intelligent analysis and decision-making layer, and collaborative control and execution layer. Each layer interacts with the others through strictly defined interfaces and data flows, forming a closed-loop intelligent management ecosystem.

[0052] The IoT sensing layer, serving as the system's foundational data source, is responsible for collecting operational status data and environmental parameters from various electromechanical terminal devices within the tunnel. This sensing layer comprises multiple sensor nodes and intelligent terminals deployed within the tunnel. These nodes and terminals are precisely divided into four monitoring units according to their functions: a structural monitoring unit, an environmental monitoring unit, a traffic flow monitoring unit, and an equipment status monitoring unit. The structural monitoring unit employs a fiber optic grating sensor array and distributed strain gauges to continuously monitor the stress, strain, and displacement changes of the tunnel lining at a fixed sampling frequency of 1 time / second.

[0053] Fiber Bragg grating sensor arrays are deployed along the tunnel wall at 5-meter intervals, with each sensor node having a built-in temperature compensation module to ensure the accuracy of strain measurement data. Distributed strain gauges are installed at critical stress points in the tunnel lining, such as the connection between the arch and the sidewalls, achieving micro-strain level measurement accuracy.

[0054] All structural monitoring data is uploaded to the data fusion processing layer in real time via industrial Ethernet. The data packet contains the sensor's unique identifier, timestamp, strain value, temperature value, and data quality identifier.

[0055] The environmental monitoring unit integrates temperature and humidity sensors, a visibility detector, and a multi-gas concentration analyzer to collect real-time micro-environmental parameters within the tunnel. The temperature and humidity sensors use digital probes, with a measurement range covering -20 degrees Celsius to 60 degrees Celsius, a relative humidity measurement range of 5% to 95%, and a data update frequency of once every 10 seconds. The visibility detector operates based on the forward scattering principle and is installed on the tunnel sidewall at a height of 3 meters above the ground, with a measurement range of 0 to 1500 meters.

[0056] The multi-gas concentration analyzer incorporates an electrochemical sensor and an infrared absorption module, enabling simultaneous detection of carbon monoxide, nitrogen oxides, sulfur dioxide, and inhalable particulate matter concentrations in tunnel air, with detection limits down to the part per million level. All sensor data from the environmental monitoring unit are aggregated via a wireless IoT communication protocol. Each data frame includes the monitoring point's geographic location code, pollutant type, concentration value, unit, and device self-test status code.

[0057] The traffic flow monitoring unit uses millimeter-wave radar and high-definition video cameras to acquire macroscopic traffic operation information within the tunnel every 5 seconds. The millimeter-wave radar is deployed at the tunnel entrance, exit, and key internal sections, employing frequency-modulated continuous wave technology to accurately detect the speed, position, and length of each vehicle, with a speed measurement error of less than 3 km / h.

[0058] The high-definition video camera is equipped with strong light suppression and fog penetration functions, with a resolution of 1920×1080 pixels and a frame rate of 25fps. The video stream data is processed through edge computing nodes, and a deep learning object detection algorithm is used to identify vehicle types, count traffic flow, and calculate average vehicle speed and vehicle density in real time.

[0059] The traffic flow data output format includes time interval, lane number, flow count, average speed, density value, and data confidence score.

[0060] The equipment status monitoring unit acquires the electrical and operational status data of key electromechanical equipment such as fans, lighting fixtures, fire pumps, and variable message signs in real time through built-in or external intelligent acquisition modules.

[0061] For wind turbine equipment, the monitoring parameters include three-phase voltage, three-phase current, active power, reactive power, power factor, winding temperature, vibration acceleration, and cumulative operating time.

[0062] Monitoring of lighting fixtures includes input voltage, operating current, lamp surface temperature, light decay coefficient, and switch status.

[0063] Monitoring fire pumps involves pump outlet pressure, motor current, bearing temperature, number of start-stop cycles, and fault alarm codes.

[0064] The monitoring of variable information flags includes the brightness uniformity of each pixel on the display screen, the status of the communication link, the content refresh success rate, and the operating voltage of the power module.

[0065] All device status data are collected through physical interfaces that conform to the industrial fieldbus standard. The sampling period is dynamically adjusted according to the device characteristics, ranging from 100ms to 10s.

[0066] Please refer to the attached document. Figure 2 The data fusion processing layer is responsible for the in-depth processing and fusion of heterogeneous data uploaded from the device IoT sensing layer. This layer consists of three core modules: a protocol adaptation module, a data cleaning module, and a feature fusion engine.

[0067] The protocol adaptation module is key to connecting diverse terminal devices to the system. It has built-in parsing libraries for at least 8 mainstream industrial fieldbus protocols and 3 wireless IoT communication protocols.

[0068] The industrial fieldbus protocol parsing library supports protocol families including ModbusTCP, PROFIbusDP, DeviceNet, CANopen, EtherCAT, PROFINETIO, CC-LinkIE, and BACnetIP. The wireless IoT communication protocol parsing library covers LoRaWAN, NB-IoT, and ZigBee 3.0 standards.

[0069] After receiving the original data packet, the protocol adaptation module first performs link layer verification and parsing, extracts the application layer data payload, and then maps the data to the unified data model defined within the system according to the pre-configured device description file.

[0070] The unified data model is defined using JavaScript object representation format. Each data point contains a globally unique device identifier, parameter type encoding, value, quality stamp, acquisition timestamp, and source protocol type.

[0071] The data cleaning module performs quality improvement processing on the protocol-adapted data stream. This module first performs outlier detection, using a statistical method based on the 3-sigma criterion to identify and mark data points that significantly deviate from the normal range.

[0072] For data marked as abnormal, the module performs removal or retention operations according to a preset strategy.

[0073] Subsequently, the data cleaning module processes the missing values ​​in the data stream, using a mean-shifting algorithm based on a sliding time window for imputation and repair.

[0074] The algorithm uses 30 valid data points before and after the current missing point to form a time window, and calculates the weighted average of the data within the window as the imputation value. The weight is inversely proportional to the time difference between the data point and the missing point.

[0075] The cleaned data stream ensures continuity and integrity. Each data packet is appended with a data quality score before output, which integrates anomaly detection results, missing value ratio, and time continuity indicators.

[0076] The feature fusion engine is the core intelligent component of the data fusion processing layer, responsible for spatiotemporal alignment and deep feature extraction of cleaned multi-source heterogeneous data.

[0077] Spatiotemporal alignment processing first relies on the high-precision BeiDou satellite positioning system and network time protocol to assign each incoming data a timestamp with nanosecond-level precision and spatial coordinates with centimeter-level precision.

[0078] The time synchronization server is deployed in the computer room of the tunnel management center and synchronizes with the time modules of each sensing unit through a fiber optic network.

[0079] Spatial coordinates are assigned based on the predefined positions of the sensors in the tunnel construction information model.

[0080] After completing the spatiotemporal alignment, the feature fusion engine uses a graph neural network-based association mining algorithm to extract the nonlinear coupling features between equipment state parameters and environmental factors and traffic load.

[0081] This graph neural network models each monitoring point in the tunnel as a node in the graph, and the edges between the nodes represent physical connections or functional associations.

[0082] The node feature vector is composed of the multi-dimensional monitoring data of that point.

[0083] Graph neural networks learn the complex interactions between nodes through multi-layer message passing and aggregation operations, and finally output an enhanced feature representation of each node. These features characterize the correlation between the local state and the global system situation.

[0084] Please refer to the attached document. Figure 3 The digital twin model layer constructs and maintains a dynamic mirror image of the physical tunnel in virtual space. This model layer consists of three intercoupled sub-models: a geometric model, a physical model, and a behavioral rule model.

[0085] The geometric model is constructed using lightweight Building Information Modeling (BIM) technology, achieving a Level 500 level of detail. It accurately recreates the internal structural outline of the tunnel, segment joints, equipment mounting bases, cable tray routing, ventilation duct layout, and fire protection pipeline distribution. The geometric model is created using parametric modeling methods, supporting rapid instantiation based on actual tunnel design drawings.

[0086] The physical model incorporates mathematical equations describing the physical laws of the tunnel, including the elastoplastic constitutive relations of the tunnel lining, the simplified form of the Navel-Stokes equations for airflow within the tunnel, and empirical models of performance degradation for key equipment such as fans and lighting fixtures.

[0087] The parameters of the physical model are calibrated using on-site measured data and historical operation and maintenance records to ensure that the error between the simulated physical quantities and the real-world observations is within 5%.

[0088] The behavioral rule model defines the interaction logic and control strategy of each entity in the tunnel system under the triggering of specific events in the form of state machines and production rules.

[0089] For example, the model clearly defines a series of linked operations that the system should perform when a fire alarm signal is triggered: first, turn off the general lighting in the area adjacent to the fire point and start the emergency lighting system; second, based on the location of the fire point and the smoke diffusion simulation results, start a specific group of jet fans to form a directional smoke exhaust airflow; at the same time, publish evacuation guidance information on variable message signs and link with the traffic signal control system to restrict vehicles from entering the tunnel.

[0090] The behavior rule model is described using Extensible Markup Language, which supports online editing and version management.

[0091] The digital twin model layer is updated through a real-time data-driven engine with microsecond-level time synchronization accuracy.

[0092] The engine takes the enhanced feature data stream output from the data fusion processing layer as input, solves the physical model and behavior rule model in real time, updates the state of all entities in the virtual tunnel, and ensures that the virtual model and the physical tunnel maintain a highly consistent state mapping at all times.

[0093] Please refer to the attached document. Figure 2 With appendix Figure 4 The intelligent analysis and decision-making layer performs advanced intelligent analysis functions based on the panoramic status information provided by the digital twin model layer.

[0094] This layer integrates three core modules: equipment health assessment module, multi-step risk prediction module, and emergency response strategy generation module.

[0095] The equipment health assessment module is responsible for quantitatively assessing the health status and predicting the lifespan of key electromechanical equipment within the tunnel.

[0096] This module uses a deep residual network to construct a model for calculating the device health index.

[0097] A deep residual network consists of one input layer, 15 hidden layers, and one output layer.

[0098] The number of nodes in the input layer corresponds to the dimension of the multidimensional feature vector extracted from the device status data. These features include time-domain features such as mean, variance, and peak factor; frequency-domain features such as spectral centroid and frequency standard deviation; and time-frequency-domain features such as wavelet packet energy entropy. The hidden layer adopts a residual block structure, with each residual block containing two weighted layers and one shortcut connection. The activation function uses a modified linear unit.

[0099] The output layer consists of one node, which uses the Sigmoid activation function to output the device's health index. This index is a continuous value between 0 and 1, with a higher value indicating a healthier device.

[0100] The deep residual network for the equipment health assessment module is trained using historical data.

[0101] The training process begins by collecting a large amount of historical operating data covering normal equipment conditions, various typical failure modes, and performance degradation processes, forming a training sample set. The number of samples is typically no less than 10,000.

[0102] The sample data is normalized so that its mean is 0 and its variance is 1.

[0103] Then, a deep residual network model is constructed, and its loss function is defined as the mean squared error function.

[0104] The network model was trained using the backpropagation algorithm and the adaptive moment estimation algorithm, with the initial learning rate set to 0.001 and an exponential decay strategy employed.

[0105] Training continues until the root mean square error of the model's predictions on the independent validation set is less than 0.05.

[0106] After training, the model is deployed to an online inference server to receive device status monitoring data in real time, calculate and output a health index.

[0107] When the health index of any device falls below the preset threshold of 0.85, the system automatically generates an early warning message and recommends preventative maintenance.

[0108] The multi-step risk prediction module uses a long short-term memory network to perform time series analysis on the fused tunnel environment, traffic flow and structural monitoring data to predict tunnel operation risks in the future.

[0109] This Long Short-Term Memory (LSTM) network consists of two stacked LSM layers and one fully connected output layer.

[0110] The input data consists of multivariate time series data within a 60-minute time window, with a series length of 720 time steps. Input feature dimensions include environmental parameters, traffic flow parameters, and key equipment status parameters.

[0111] The first long short-term memory layer has 64 hidden units, and the second long short-term memory layer has 32 hidden units.

[0112] The fully connected output layer has 6 nodes, corresponding to risk probability predictions for the next 6 time steps. Each time step is 5 minutes apart, so the total prediction time is 30 minutes.

[0113] The network output is normalized using the Softmax function, representing the probability of a specific risk event occurring at each future time step.

[0114] The training of this long short-term memory network uses weighted cross-entropy as the loss function, which assigns higher weights to high-risk samples to improve the model's predictive sensitivity for minority dangerous events.

[0115] The weighting is inversely proportional to the frequency of risk events in the sample.

[0116] Training data comes from historical operational records and accident reports, and must contain a sufficient number of positive samples to ensure the model's generalization ability.

[0117] When the probability of any risk event predicted by the model within the next 30 minutes is greater than 70%, the system will activate a high-level warning and notify operations personnel to intervene.

[0118] The emergency response strategy generation module is based on reinforcement learning algorithms and performs strategy optimization in a virtual environment constructed by a digital twin model.

[0119] This module models the tunnel emergency response environment as a Markov decision process.

[0120] The state space S consists of the tunnel panorama state variables output in real time by the digital twin model, and its dimensions can reach hundreds of dimensions.

[0121] Action space A corresponds to all executable control command combinations, including discrete actions such as fan start / stop, lighting mode switching, variable message sign content update, and traffic signal control.

[0122] reward function It is a comprehensive scalar function whose calculation takes into account the effectiveness of accident control, rescue time, impact on normal traffic, and resource consumption costs.

[0123] Specifically, the reward function can be expressed as:

[0124] ;

[0125] Among them, the accident control index quantifies the degree to which the accident is controlled according to its type; the rescue time is the time elapsed from the start of the decision-making process to the stabilization of the situation; the traffic impact index measures the traffic delays caused by the response measures; and the resource cost calculation includes electricity consumption, equipment wear and tear, etc. to These are pre-set positive weighting coefficients.

[0126] The Markov decision process is solved using a proximal policy optimization algorithm. The agent, i.e., the policy network, adopts an actor-critic architecture. The actor network is responsible for outputting the probability distribution of actions based on the current state, while the critic network is responsible for evaluating the value of the current state.

[0127] The agent interacts with the digital twin environment, collecting data tuples of state, action, reward, and next state over a large number of simulation training rounds, and using this data to update the parameters of the policy network and value network.

[0128] The training objective is to optimize the expected cumulative discount reward. After sufficient training, the policy network can quickly generate efficient response strategies for various initial accident states fed back by the digital twin model.

[0129] When a sudden event occurs in a real tunnel, the module inputs the current real state into the trained policy network, and the network outputs a recommended optimal control action sequence, which is immediately sent to the collaborative control execution layer.

[0130] The collaborative control execution layer is the final link in the implementation of system decisions, responsible for safely and reliably executing the control commands generated by the intelligent analysis and decision-making layer in the physical tunnel.

[0131] This layer communicates with the execution terminals at the tunnel site through a distributed control network.

[0132] The distributed control network adopts a hybrid networking approach of industrial Ethernet and wireless IoT. Core control nodes are interconnected via gigabit fiber optic cables, while edge execution nodes are accessed wirelessly via 5G or Wi-Fi 6.

[0133] The collaborative control execution layer has strict instruction verification and security interlocking functions. The instruction verification module performs syntax checks, semantic analysis, and permission verification on each received control instruction.

[0134] Syntax checking ensures that the command format conforms to the predefined command specifications; semantic analysis verifies that the command parameters are within the operating range allowed by the device; permission verification confirms the legality of the command source and the operating permission level required to execute the command.

[0135] Safety interlocking is a critical mechanism for ensuring the safe operation of a system. It maintains a complex set of logical rules that define the interlocking relationships and safety constraints between devices.

[0136] For example, after the system receives an instruction to start a group of fans for longitudinal ventilation, the safety interlock module will simultaneously check the status of fire detectors in relevant areas of the tunnel, vehicle location information detected by traffic flow radar, and carbon monoxide concentration monitoring values.

[0137] Only when all safety preconditions are met—namely, the fire alarm signal has been verified as genuine and valid through multiple verifications, there are no vehicles obstructing the air duct path, and the concentration of toxic gases in critical areas does not exceed the limit—will the safety interlock module send the final start command to the designated fan controller.

[0138] If any precondition is not met, the instruction will be suspended and an alarm log will be generated, notifying the operations staff to intervene manually.

[0139] The collaborative control execution layer also supports the ability to dynamically adjust control strategies. This capability is evaluated based on real-time control performance feedback from a digital twin model.

[0140] The system continuously compares the difference between the actual readings of the sensors in the physical tunnel and the ideal state predicted by the digital twin model after the control commands are executed.

[0141] When the deviation exceeds a preset threshold, the dynamic adjustment algorithm is activated. This algorithm uses the incremental proportional-integral-derivative control principle to fine-tune the parameters of the control command.

[0142] For example, in wind turbine speed control, if the actual wind speed is less than the model's expectation, the algorithm will increase the wind turbine speed setpoint by a certain step size until the actual wind speed enters the target range.

[0143] This closed-loop optimization mechanism ensures the accuracy of control actions and optimizes the overall operating efficiency and stability of the system.

[0144] The system described in this embodiment achieves closed-loop management of the entire process of tunnel physical entity and virtual model through the precise collaboration of the above five levels.

[0145] From the underlying multi-source data perception, to the mid-level fusion processing and twin modeling, and then to the upper-level intelligent decision-making and precise execution, the system forms a highly autonomous and continuously optimized intelligent management entity.

[0146] Its value lies in improving tunnel operation safety, maintenance efficiency, and total life cycle cost.

[0147] When deploying the system, a central server cluster needs to be built in the tunnel management center to carry the core computing tasks of the data fusion processing layer, digital twin model layer and intelligent analysis and decision-making layer.

[0148] At the tunnel site, corresponding sensing and control network infrastructure needs to be deployed according to the aforementioned design, and interconnected with the central system through high-bandwidth, low-latency communication links. The system's software platform adopts a microservice architecture, allowing each functional module to be developed, deployed, and expanded independently, and communicating through well-defined application programming interfaces, ensuring the system's maintainability and technological evolution capabilities.

[0149] This embodiment also provides another implementation of a tunnel management system based on digital twins, the core feature of which is an alternative technical solution for the equipment health assessment module in the intelligent analysis and decision-making layer.

[0150] In this embodiment, the equipment health assessment module does not use deep residual networks, but is based on an ensemble learning framework, specifically the gradient boosting decision tree algorithm, to model and predict the equipment health index.

[0151] This approach is suitable for application scenarios with relatively low data feature dimensions but a large number of samples, and it also has the advantage of strong model interpretability.

[0152] The gradient boosting decision tree model consists of hundreds of weak classification tree sequences. The maximum depth of each decision tree is limited to 6 layers to prevent overfitting.

[0153] The input features of the model are also time-domain, frequency-domain and time-frequency-domain features extracted from equipment status monitoring data, but the feature dimensions are usually controlled within 50 dimensions.

[0154] The training process employs a phased additive model strategy.

[0155] In the initial stage, the first decision tree uses all the training data to fit the residuals of the device health index.

[0156] In each subsequent stage, the newly added decision tree is dedicated to fitting the residual predictions of the model from the previous stage.

[0157] The learning rate was set to 0.1, and the subsampling ratio was set to 0.8.

[0158] The training stops when the performance on the validation set no longer improves over 20 consecutive iterations.

[0159] After the gradient boosting decision tree model of the equipment health assessment module is trained, its inference process is as follows: For the real-time collected equipment status data, feature extraction and standardization are performed in the same way as in the training phase. Then, the feature vectors are input into the trained gradient boosting decision tree model.

[0160] The model outputs a device health index between 0 and 1 by weighted summation of the predictions from all component trees.

[0161] This index is also used to trigger alerts, with a threshold set at 0.85.

[0162] Gradient boosting decision tree models provide a ranking of feature importance, helping maintenance personnel understand the key factors affecting equipment health. For example, vibration features may contribute more to the health assessment of a certain type of fan than temperature features, thus providing a basis for targeted maintenance activities.

[0163] The specific implementation methods of the remaining system components in this embodiment, including the device IoT sensing layer, data fusion processing layer, digital twin model layer, multi-step risk prediction module, emergency response strategy generation module, and collaborative control execution layer, are all consistent.

[0164] The system architecture, data flow, control logic, and security mechanisms all follow the same design principles and technical specifications.

[0165] This alternative enriches the system's technology options, enabling it to adapt to different data foundations and operational needs.

[0166] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0167] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tunnel management system based on digital twins, characterized in that, include: The equipment IoT sensing layer is used to collect the operating status data and environmental parameters of various electromechanical terminal equipment in the tunnel. The equipment IoT sensing layer includes a variety of sensor nodes and smart terminals deployed in the tunnel. These nodes and terminals are divided into structural monitoring units, environmental monitoring units, traffic flow monitoring units and equipment status monitoring units according to their functions. The data fusion processing layer is used to parse, clean and standardize heterogeneous data from different protocol interfaces, and to complete the spatiotemporal alignment and feature extraction of multi-source data. The data fusion processing layer includes a protocol adaptation module, a data cleaning module and a feature fusion engine. The digital twin model layer is used to dynamically drive the virtual tunnel model based on the processed data, so as to realize the real-time mirroring of the physical entity in the virtual space. The digital twin model layer consists of a geometric model, a physical model and a behavior rule model. The intelligent analysis and decision-making layer is used to perform equipment health assessment, risk prediction and emergency response strategy generation based on the panoramic status information output by the digital twin model. The intelligent analysis and decision-making layer integrates the equipment health assessment module, the multi-step risk prediction module and the emergency response strategy generation module. The collaborative control execution layer is used to implement precise linkage control of relevant equipment groups in the tunnel according to decision instructions. The collaborative control execution layer communicates with the execution terminal at the tunnel site through a distributed control network and has instruction verification and safety interlocking functions.

2. The tunnel management system based on digital twin according to claim 1, characterized in that, The structural monitoring unit uses a fiber optic grating sensor array and distributed strain gauges to continuously monitor the stress, strain and displacement changes of the tunnel lining at the same sampling frequency. The environmental monitoring unit integrates a temperature and humidity sensor, a visibility detector, and a multi-gas concentration analyzer to collect real-time data on temperature, humidity, visibility, and carbon monoxide and nitrogen oxide concentrations inside the tunnel. The traffic flow monitoring unit uses millimeter-wave radar and high-definition video cameras to acquire information on traffic flow, average vehicle speed, and vehicle density within the tunnel at fixed intervals. The equipment status monitoring unit acquires the voltage, current, power factor, running time, and fault code status of key electromechanical equipment such as fans, lighting fixtures, fire pumps, and variable information signs in real time through built-in or external intelligent acquisition modules.

3. A tunnel management system based on digital twins according to claim 1, characterized in that, The protocol adaptation module has a built-in parsing library for multiple industrial fieldbus protocols and multiple wireless IoT communication protocols, which can automatically identify the communication protocols of access devices and convert their data into the standard data format inside the system. The data cleaning module detects and removes outliers from the raw data stream and uses a mean shift algorithm based on a sliding time window to imput and repair missing data, ensuring the continuity and integrity of the data stream. The feature fusion engine performs spatiotemporal alignment processing on the cleaned multi-source heterogeneous data, assigns a unified timestamp and spatial coordinates to each data point through the high-precision BeiDou satellite positioning system and network time protocol, and uses a graph neural network-based association mining algorithm to extract deep coupling features between equipment status parameters and environmental factors and traffic load.

4. A tunnel management system based on digital twins according to claim 1, characterized in that, The geometric model is constructed using lightweight building information modeling technology to accurately reproduce the internal structure of the tunnel, equipment layout, and pipeline routing. The physical model incorporates the mechanical constitutive relationship of the tunnel lining, the aerodynamic equations, and the aging and wear patterns of the equipment. It is used to simulate the mechanical response of the tunnel structure under vehicle loads and environmental effects, as well as the degradation trend of equipment performance over time. The behavioral rule model defines the interaction logic and control strategy between entities within the tunnel system under the triggering of specific events; The digital twin model layer ensures dynamic consistency between the virtual model and the physical tunnel state through a real-time data-driven engine with microsecond-level time synchronization accuracy.

5. A tunnel management system based on digital twins according to claim 1, characterized in that, The equipment health assessment module uses real-time data and historical operation records uploaded by the equipment status monitoring unit to construct an equipment health index calculation model using a deep residual network. This model can identify early signs of equipment performance degradation and predict its remaining service life. When the health index is less than a preset threshold, it will automatically issue an early warning. The multi-step risk prediction module uses a long short-term memory network to perform time series analysis on the fused tunnel environment, traffic flow and structure monitoring data. The emergency response strategy generation module is based on reinforcement learning algorithm. It performs parallel simulation and deduction of multiple potential emergency response plans in a digital twin model, evaluates the comprehensive utility of each plan in terms of execution effect, resource consumption and secondary risks, and finally selects the strategy with the highest utility value as the recommended plan output.

6. A tunnel management system based on digital twins according to claim 5, characterized in that, The training process of the deep residual network in the device health assessment module is as follows: A large amount of historical operating data covering normal equipment conditions, various typical failure modes, and performance degradation processes was collected to form a training sample set; The sample data is normalized and multi-dimensional feature vectors, including time-domain features, frequency-domain features, and time-frequency-domain features, are extracted. A deep residual network model with hidden layers is constructed, where the number of input layer nodes is equal to the dimension of the feature vectors, and the output layer corresponds to the health index of the device. The network model was trained using the backpropagation algorithm and the adaptive moment estimation algorithm.

7. A tunnel management system based on digital twins according to claim 5, characterized in that, The construction and operation process of the long short-term memory network in the multi-step risk prediction module is as follows: The network packet consists of a long short-term memory layer and a fully connected output layer; the input data is a sequence of tunnel environment data, traffic flow data, and key equipment status data within a past time window, which has been preprocessed by the data fusion processing layer. The network learns long-term dependencies in historical sequences and outputs a probability sequence of specific risk events occurring within the tunnel at future time steps; The loss function used to train this network is weighted cross-entropy, which assigns higher weights to high-risk samples.

8. A tunnel management system based on digital twins according to claim 5, characterized in that, The specific implementation framework of the reinforcement learning algorithm in the emergency response strategy generation module is as follows: The tunnel emergency response environment is modeled as a Markov decision process, where the state space consists of the tunnel panoramic state variables output in real time by the digital twin model, the action space corresponds to all executable control command combinations, and the reward function comprehensively considers the accident control effect, rescue time, impact on normal traffic, and resource consumption cost. The Markov decision process is solved using a proximal policy optimization algorithm. The agent learns the optimal policy network for generating efficient disposal strategies through interactive simulation training with a digital twin environment.

9. A tunnel management system based on digital twins according to claim 1, characterized in that, The instruction verification and security interlocking functions of the collaborative control execution layer specifically include: The system performs syntax checks, semantic analysis, and permission verification on the received control commands. Upon receiving a command to start the fans for longitudinal ventilation, it simultaneously verifies the status of fire detectors and traffic flow in the tunnel. Only when it is confirmed that there are no vehicles blocking the ventilation ducts and that the fire alarm signal is genuine and valid will the system send a start command to the designated fan group.

10. A tunnel management method based on digital twins, characterized in that, The digital twin-based tunnel management system described in any one of claims 1-9 is used to implement digital twin-based tunnel management.

Images