A tunnel model intelligent carrying and safety state visualized deduction system and method

Abstract

The application discloses a tunnel model intelligent carrying and safety state visualized deduction system and method, which comprises a multidimensional state inversion and early warning module, a moving vehicle and an adjustable bearing platform. The multidimensional state inversion and early warning module comprises a physical perception layer, a data driving layer, an algorithm decision layer and a visualized deduction layer, wherein the algorithm decision layer is embedded with a stress and displacement prediction model based on a ConvLSTM architecture, can inversely calculate the global structure state of the moving vehicle according to real-time sensing data and generate an early warning signal. The moving vehicle has a jacking and rotating function and can realize heavy load rail replacement in a limited space. The adjustable bearing platform is composed of honeycomb structure units and can be flexibly adjusted according to the size of the tunnel model. The visualized deduction layer maps the inversion result to a three-dimensional grid model in real time and generates a dynamic stress and displacement cloud chart. The application realizes real-time inversion of the structure state in the heavy load transportation process, intelligent rail replacement and visualized early warning.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering testing technology, and in particular to a system and method for intelligent transportation and safety status visualization simulation of tunnel models. Background Technology

[0002] Tunnels are relatively enclosed spaces, and tunnel linings, as the main load-bearing structures, are highly susceptible to concrete cracking, component connection failure, and even complete collapse under high-temperature environments. Therefore, simulating the thermal coupling damage evolution of tunnel linings under real fire conditions through large-scale experiments is of great significance for tunnel disaster prevention and mitigation.

[0003] Currently, research on the thermo-coupling performance of tunnel structures, both domestically and internationally, largely focuses on small components or uses horizontally arranged full-ring tests. These tests typically employ uniform temperature fields for heating, making it difficult to simulate the lining damage caused by non-uniform temperature fields in real fires. In light of this, Chinese Patent Application No. 2021102117560 discloses a multi-dimensional spatial loading fire test system for tunnel structures and its implementation method. This prior art technology achieves full-process simulation of thermo-coupling damage in tunnel structures. In such large-scale vertical test systems, the tunnel model typically possesses physical characteristics of large size (several meters high) and heavy load (tens or even hundreds of tons). However, in actual test operations, limited by the spatial layout of the test site, the tunnel model faces the following technical bottlenecks in transportation, track changing, and positioning:

[0004] 1. Structural health monitoring suffers from locality and lag: When a heavy-load tunnel model is moved within a confined space, the shift of the center of gravity or uneven local stress can easily lead to the overall instability of the transport structure. Traditional transport vehicle monitoring relies heavily on discretely distributed sensors for point-based numerical data acquisition. This conventional monitoring method can only reflect the single mechanical response of local measuring points and is difficult to reconstruct the overall stress and displacement distribution of the transport vehicle frame under complex conditions such as eccentric compression and dynamic jacking and track replacement. Furthermore, due to the lack of a systematic overall stress state mapping mechanism, the assessment of structural damage and instability risks is delayed, making it difficult to achieve real-time dynamic early warning under heavy-load extreme conditions.

[0005] 2. Poor adaptability to heavy-load track switching under space-constrained conditions: Due to the layout of the test site, tunnel models often need to be reversed or track switched during transportation to the loading position. Existing transport equipment mostly uses fixed chassis or conventional unidirectional travel mechanisms. When carrying models weighing tens of tons, performing track switching and turning operations in a limited space can easily cause distortion of the equipment chassis structure or stress concentration, making it difficult to balance the structural safety of heavy-load transportation with the efficiency of cyclical testing.

[0006] 3. Lack of adjustment mechanisms for loading height under large-span loads: The dimensions of tunnel test segments vary significantly between different batches, while the stroke of the circumferential loading equipment in the test system is limited. For segments with smaller dimensions, traditional methods often rely on temporarily assembled rigid blocks for auxiliary support. There is a lack of an adaptive platform that can meet heavy load requirements and allow for modular height adjustment based on the segment's geometry. Furthermore, traditional support adjustment methods are cumbersome and prone to local pressure instability under heavy loads and high-pressure conditions.

[0007] To address the aforementioned technical bottlenecks, there is an urgent need to develop a tunnel model intelligent transportation and safety status visualization simulation system and its implementation method, possessing real-time inversion of the entire mechanical state, intelligent track switching, and height adaptive adjustment functions. By achieving quantitative assessment and digital dynamic early warning of structural safety during heavy-load transportation, this system will provide reliable equipment and technical support for large-scale tunnel structure thermo-mechanical coupling tests. Summary of the Invention

[0008] To address the aforementioned technical shortcomings, the present invention aims to provide a system and method for intelligent transportation and safety status visualization of tunnel models. Through the deep integration of collaborative design of physical structures and intelligent algorithms, this system solves the problems of automatic track changing and real-time inversion of equipment health status during tunnel model transportation. Based on preset commands, the system can automatically change tracks during operation and perform visual monitoring of structural health status, providing technical support for the intelligent transformation and full lifecycle monitoring of equipment.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] This invention provides a tunnel model intelligent transportation and safety status visualization simulation system, comprising:

[0011] A multi-dimensional state inversion and early warning module, a transport vehicle, and an adjustable support platform mounted on the transport vehicle; wherein, the multi-dimensional state inversion and early warning module includes:

[0012] The physical sensing layer is used to collect mechanical and attitude data of the transport vehicle in real time during its operation.

[0013] The data-driven layer is used to preprocess the raw data collected by the physical sensing layer and transmit it to the algorithm decision layer.

[0014] The algorithm decision layer embeds a stress and displacement prediction model based on the ConvLSTM (Convolutional Long Short-Term Memory Network) architecture. The stress and displacement prediction model is used to invert the structural state of the transport vehicle during the transportation and track changing process in real time based on the preprocessed data, and execute the structural state judgment logic based on the inversion results to generate early warning signals.

[0015] The visualization and inference layer is used to retrieve the three-dimensional mesh model of the transport vehicle and generate dynamic visualization stress and displacement cloud maps based on the inversion results of the algorithm decision layer.

[0016] The transport vehicle includes a load-bearing frame system, a lifting and rotating system, and a moving system, and is used for transporting and changing the track of the tunnel model;

[0017] The adjustable bearing platform is composed of multiple units with a honeycomb internal structure, and the number of unit combinations and stacking layers are adaptively adjusted according to the size of the tunnel model to be transported.

[0018] Preferably, the physical sensing layer includes strain gauges, displacement gauges, velocity sensors, load sensors, and position sensors deployed at multiple preset monitoring points on the transport vehicle and the adjustable bearing platform, for collecting stress and displacement data at the monitoring points of the bearing frame system, the travel speed of the transport vehicle, the lifting speed, lifting force, rotation speed of the jacking and rotating system, and the total weight of the transport vehicle and the adjustable bearing platform.

[0019] The visualization simulation layer uses node assignment, vertex rendering, and color gradient mapping techniques to map the stress and displacement prediction data output by the algorithm decision layer to the three-dimensional mesh model of the transport vehicle in real time, generating a dynamic visualization cloud map.

[0020] Preferably, the structural state inversion logic executed by the multidimensional state inversion and early warning module includes:

[0021] S1, Based on the finite element analysis method, the structural mechanical performance dataset of the transport vehicle is obtained through preprocessing and dataset construction;

[0022] S2, using the structural mechanics performance dataset, train a stress and displacement prediction model based on the ConvLSTM architecture, determine the online inversion calculation logic of the stress and displacement prediction model, and extract the spatial topological features and temporal evolution law of the transport vehicle.

[0023] S3, the stress and displacement prediction model is trained offline using the structural mechanical performance dataset, and error verification and calibration are performed based on physical test data to ensure that the prediction error range is ≤5%;

[0024] S4. The trained stress and displacement prediction model is deployed to the multi-dimensional state inversion and early warning module, and three-dimensional mesh reconstruction and dynamic cloud map rendering are performed through the visualization inference layer.

[0025] Preferably, step S1 further includes:

[0026] S101. A finite element model is established based on the geometric dimensions, material properties, and connection relationships of the transport vehicle. After performing mesh sensitivity analysis, the model is compared and verified with physical test data.

[0027] S102 uses experimental design methods to extract multiple simulation conditions within the range of factor level values ​​covering load magnitude, bearing location, and adjustable bearing platform combination mode;

[0028] S103 uses automated scripts to perform batch solving and data extraction for multiple working conditions. It employs low-pass filtering to remove high-frequency noise, eliminates singular and invalid data, and fills in missing time series data with neighboring values.

[0029] S104 standardizes and maps the processed displacement and stress tensors, integrating them to form a time series dataset of nodal displacement and stress under various working conditions.

[0030] Preferably, the online inversion calculation logic of the stress and displacement prediction model includes:

[0031] S201, the processed real-time sensing data, namely the 5-dimensional input vector of the tunnel model weight, the adjustable bearing platform weight, and the current center of gravity coordinates, is reconstructed into a 3D spatial input tensor using a multilayer perceptron and a spatial broadcast mechanism. To accurately characterize the initial spatial distribution of the current load on the transport vehicle;

[0032] S202, input tensor The hidden state tensor preserved at the previous time step With cellular memory tensor The ConvLSTM is synchronously fed in, and the spatial force transmission path is extracted by introducing a two-dimensional spatial convolution operator.

[0033] Preferably, the stress and displacement prediction model further includes a parallel dual-branch decoding network and a joint loss function; specifically including:

[0034] S203, Physical field decoding and reconstruction, obtaining the deep hidden state of the ConvLSTM output. The vector is then fed into a parallel dual-branch decoder; the stress branch outputs a global prediction vector through transposed convolution upsampling and flattening mapping; the displacement branch outputs a global prediction vector through subpixel convolution and flattening mapping.

[0035] S204 introduces a joint smoothing L1 loss function to predict stress values ​​at the target time step. Compared with the true value Its stress smoothing loss function Defined as:

[0036] when hour, ;

[0037] when hour, ;

[0038] Similarly, the displacement branch smoothing loss function Based on threshold Segmented calculation, constructing a system containing dynamic weighting factors and Multi-task joint loss function .

[0039] Preferably, in step S4, the 3D mesh reconstruction and dynamic rendering of the cloud map includes:

[0040] S401, extract the node coordinates and element indexes of the finite element model pre-established based on the geometric dimensions, material properties and connection relationships of the transport vehicle, divide the quadrilateral elements in the finite element model into triangular patches, and generate a three-dimensional mesh model with the same geometric dimensions as the finite element model.

[0041] S402, normalize the prediction data of all grid vertices at the current time step;

[0042] S403 uses a barycentric coordinate interpolation algorithm to perform color rendering on pixels inside the surface. The interpolation calculation model is defined as follows:

[0043] ;

[0044] In the formula, , , The normalized predicted values ​​of the face vertices are given, and the weighting coefficients satisfy the following conditions: ;

[0045] S404 constructs a modular, visual user monitoring interface and integrates a trained stress and displacement prediction model based on the ConvLSTM architecture.

[0046] S405 uses the preprocessing interface of the data-driven layer to display the real-time acquired physical sensor data in the form of dynamic line graphs and statistical tables on the interface.

[0047] S406, invoke the stress and displacement prediction model to perform forward inference, and output the predicted stress and displacement values ​​of the transport vehicle node;

[0048] S407, in the monitoring interface, synchronously displays a dynamic visualization cloud map generated based on the S403 steps in the form of a 3D cloud map.

[0049] Preferably, the load-bearing frame system includes two sets of symmetrically arranged variable cross-section fish-belly type load-bearing main beams, which are connected by longitudinal beams and several secondary beams and are equipped with reinforcing nodes.

[0050] The lifting and rotating system includes several hydraulic lifting mechanisms installed at the reinforcement nodes. The hydraulic lifting mechanisms are connected to the variable cross-section fish-belly type main beam through pressure-bearing guide cylinders. The output ends of the hydraulic lifting mechanisms on the same side are connected together through anti-slip heads. The bottom surface of the anti-slip heads is in contact with the external track surface.

[0051] The mobile system includes a power supply component, a controller, warning lights, and a positioning component. The power supply component, warning lights, positioning component, lifting and rotating system, and multi-dimensional state inversion and early warning module are electrically or communicatively connected to the controller. The positioning component is used to identify the real-time coordinates and track-changing nodes of the transport vehicle and send the position signal to the controller. The controller is used to receive the early warning signal generated by the algorithm decision layer and the signal from the positioning component, and accordingly control the warning lights to perform audible and visual alarms, and issue track-changing execution commands to the lifting and rotating system.

[0052] The adjustable bearing platform is a honeycomb stacked structure, comprising several units arranged in a stepped stacked manner. Adjacent units on the same layer are horizontally locked together by T-shaped splicing grooves and L-shaped right-angle connecting seats, while units stacked on different layers are vertically positioned and locked together by positioning components.

[0053] This invention also provides a method for implementing intelligent transportation of tunnel models, comprising the following steps:

[0054] Step 1: Deploy sensors of the physical sensing layer on the transport vehicle and initialize the stress and displacement prediction model parameters of the multidimensional state inversion and early warning module.

[0055] Step 2: Assemble the adjustable bearing platform according to the size of the tunnel model to be transported, and adjust the assembled adjustable bearing platform to the center of the transport vehicle. Hoist the tunnel model onto the segment transport support installed on the adjustable bearing platform, and simultaneously update the stress boundary conditions of the current working condition to the algorithm decision layer.

[0056] Step 3: Start the transport vehicle and make it move along the preset trajectory. The algorithm decision layer synchronously inverts the stress and displacement evolution state of the entire domain nodes of the transport vehicle based on the real-time sensor data stream, and issues an early warning when the predicted state exceeds the safety threshold.

[0057] Step four: After the transport vehicle moves to the rail-changing position, it performs lifting and turning operations through the jacking and rotating system;

[0058] Step 5: After the turning is completed, the lifting and rotating system performs a reset operation, allowing the transport vehicle to fall onto the cross track and continue moving.

[0059] Step Six: In the entire transportation and track changing process described above, the sensor measured data and the full-field predicted values ​​output by the stress and displacement prediction model are mapped in real time onto the three-dimensional mesh model of the transport vehicle through the visualization simulation layer, generating a dynamic and visualized stress and displacement cloud map.

[0060] Beneficial effects:

[0061] 1. Traditional transport vehicles rely on discrete sensors for point-to-point data acquisition, which can only reflect the single mechanical response of local measurement points and is difficult to reconstruct the global stress / displacement distribution under complex working conditions such as heavy load, eccentricity, and track replacement. This invention uses a visualization and inference layer to map the one-dimensional time-series signal acquired by the physical sensing layer to a three-dimensional mesh model in real time. By using node assignment, vertex rendering, and color gradient mapping technology, dynamic cloud maps are generated, allowing maintenance personnel to intuitively grasp the overall mechanical evolution of the structure and potential risk areas.

[0062] The algorithm decision layer executes structural state judgment logic based on the inversion results output by the stress and displacement prediction model. When the predicted stress or displacement exceeds the safety threshold, an early warning signal is generated immediately, overcoming the shortcomings of the traditional method of "local monitoring - delayed judgment" and providing a reliable basis for active safety control under heavy load extreme conditions.

[0063] 2. Traditional finite element analysis is time-consuming (usually several hours) when dealing with the coupling of multiple parameters and operating conditions of transport vehicles, failing to meet the real-time control requirements on site. This invention introduces a stress and displacement prediction model based on ConvLSTM, encoding the physical laws of finite element analysis into the weights of a neural network. According to the data in Table 1, this stress and displacement prediction model achieves an average displacement R² of 0.921 and an average stress R² of 0.932 on the test set, with a mean absolute error (MAE) of only 0.983 for stress and only 0.041 for displacement. Furthermore, online forward inference after offline training takes only a few seconds, preserving the high spatial accuracy of the finite element level while meeting the real-time monitoring requirements during transportation and track changing.

[0064] To address the highly spatial nonlinearity and time-varying nature of the stress state during track switching, the algorithm's decision layer utilizes a two-dimensional convolution operator of ConvLSTM to extract the spatial force transmission path of the load within the supporting frame, while simultaneously capturing the structural stress evolution through a gating mechanism. Compared to traditional time-series prediction models (such as LSTM and GRU), this significantly reduces prediction hysteresis and abrupt errors during load condition switching.

[0065] 3. The transport vehicle adopts a variable cross-section fish-belly type load-bearing main beam and a differentiated cross-section skeleton design, combined with the reinforcement nodes of the lifting and rotating system, to ensure that the chassis is subjected to uniform stress during heavy-load track replacement, and realize automatic track replacement of tens of tons of tunnel models under space-constrained conditions.

[0066] 4. The adjustable bearing platform is composed of multiple units with a honeycomb internal structure. Units on the same layer are horizontally locked together by T-slots, I-slots, and L-shaped right-angle connectors, while units on different layers are vertically stacked by positioning components. The number of units and layers can be flexibly adjusted according to the size of the tunnel model, eliminating the cumbersome operation method of traditional temporary assembly of rigid pads.

[0067] The honeycomb internal structure achieves vehicle weight reduction while ensuring compressive stiffness, significantly reduces the torque load of the track-switching drive system, and avoids the safety hazard of local pressure instability under heavy load and high pressure conditions.

[0068] 5. The physical perception layer collects multimodal data such as the weight of the tunnel model, the weight of the adjustable load-bearing platform, the coordinates of the center of gravity, the travel speed, the lifting force, and the rotation speed. The algorithm decision layer uses a multilayer perceptron and a spatial broadcasting mechanism to reshape the low-dimensional input into a three-dimensional spatial input tensor, accurately representing the initial spatial distribution of the load on the transport vehicle.

[0069] A parallel dual-branch decoding network is employed to output global stress and displacement prediction vectors separately. An innovative joint smoothing L1 loss function is introduced, which adopts the L2 norm when the prediction error is less than a threshold (ensuring rapid convergence under small fluctuations) and the L1 norm when the error exceeds the threshold (due to shocks / track changes) (effectively truncating amplified abnormal gradients). Simultaneously, a dynamic weighting factor balances the gradient dominance of stress and displacement. This mechanism achieves an optimal balance between global smoothness and robustness in capturing local singularities in the stress and displacement prediction model.

[0070] 6. The visualization and simulation layer extracts the node coordinates and element indices of the finite element model, divides the quadrilateral element into triangular patches, and uses a centroid coordinate interpolation algorithm to continuously color render the pixels inside the patches. This allows for high-frame-rate display of the mechanical response state of the transport vehicle on the intelligent control terminal without relying on bulky finite element post-processing software.

[0071] The system synchronously records sensor measured data and model predicted values ​​throughout the entire transportation and track changing process, and displays them in a visualization interface in the form of dynamic line graphs, statistical tables and 3D cloud maps, providing reliable data support for equipment life cycle health monitoring, accident backtracking and subsequent design optimization. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1This is a schematic diagram of an intelligent track-changing and transporting system for tunnel models provided in an embodiment of the present invention.

[0074] Figure 2 This is a schematic diagram of the transport vehicle structure provided in an embodiment of the present invention;

[0075] Figure 3 This is a schematic diagram of the steel frame of the transport vehicle provided in an embodiment of the present invention;

[0076] Figure 4 A schematic diagram of the adjustable load-bearing platform structure provided in an embodiment of the present invention (connection relationship between the L-shaped right-angle connector and the unit).

[0077] Figure 5 This is a schematic diagram of the adjustable load-bearing platform structure provided in an embodiment of the present invention;

[0078] Figure 6 This is a flowchart of the structural state inversion method for the multidimensional state inversion and early warning module provided in an embodiment of the present invention;

[0079] Figure 7 A diagram of a stress and displacement prediction model based on a ConvLSTM architecture provided for an embodiment of the present invention;

[0080] Figure 8 A scatter regression diagram of the stress prediction and actual values ​​of the stress and displacement prediction model provided in the embodiment of the present invention;

[0081] Figure 9 A scatter regression diagram of the displacement prediction value and the actual value of the stress and displacement prediction model provided in the embodiment of the present invention;

[0082] Figure 10 A flowchart illustrating the implementation method of the intelligent track-changing transportation system for tunnel models provided in this embodiment of the invention;

[0083] Figure 11 This is a schematic diagram of the position of the transport vehicle before transportation, provided in an embodiment of the present invention.

[0084] Figure 12 This is a schematic diagram showing the location of the transport vehicle after transportation, provided in an embodiment of the present invention.

[0085] Figure 13 This is a schematic diagram of the structure of the L-shaped right-angle connector provided in an embodiment of the present invention;

[0086] Figure 14 This is a schematic diagram of the structure of the segment transport support provided in an embodiment of the present invention.

[0087] Figure label:

[0088] 1-Transfer vehicle; 2-Adjustable load-bearing platform; 3-Steel frame; 4-Protective enclosure; 5-Variable cross-section fish-belly type main load-bearing beam; 6-Longitudinal beam; 7-Universal wheel set; 8-Reinforcement node; 9-Secondary beam; 10-Auxiliary beam; 11-Power supply assembly; 12-Warning light; 13-Positioning assembly; 14-L-shaped right-angle connector; 15-Positioning pin; 16-Fastening bolt; 17-Hydraulic lifting mechanism; 18-Pressure-bearing guide cylinder; 19-Anti-slip top; 20-I-shaped groove; 21-Tunnel model; 22-Segment transport support. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] This invention provides an intelligent transportation and safety status visualization simulation system for tunnel models. The system mainly consists of a hardware execution architecture and a software decision-making architecture, including an intelligent track-changing transportation system for tunnel models and a multi-dimensional status inversion and early warning module. Wherein:

[0091] The intelligent track-changing transport system for the tunnel model is used to directly carry and transport the tunnel model 21. Specifically, it consists of a transport vehicle 1 and an adjustable support platform 2 mounted on the transport vehicle 1. The multi-dimensional state inversion and early warning module is communicatively connected to the intelligent track-changing transport system for the tunnel model, used for visual monitoring of the entire transport process. The transport vehicle 1 includes a support frame system, a lifting and rotating system, and a moving system. The adjustable support platform 2 is mounted on the transport vehicle 1 and consists of multiple units with a honeycomb internal structure. These units are laterally locked through T-slots, I-slots 20, and L-shaped right-angle connecting seats 14, and vertically stacked through positioning pins 15 and fastening bolts 16. The number of units and layers can be flexibly adjusted according to the segment size. (See L-shaped right-angle connecting seat for details.) Figure 13 ;

[0092] See Figures 2-3 The load-bearing frame system consists of a steel frame 3 and a protective enclosure 4 laid on top of it. The steel frame 3 includes two sets of symmetrically arranged variable cross-section fish-belly type main load-bearing beams 5. The mid-span of the two sets of variable cross-section fish-belly type main load-bearing beams 5 is connected by longitudinal beams 6. The bottom of the steel frame 3 is provided with four universal wheel sets 7. The steel frame 3 is provided with reinforcing nodes 8 and secondary beams 9 at the connection positions with the universal wheel sets 7. An auxiliary beam 10 is provided between the longitudinal beams 6 and the secondary beams 9. The cross-sectional height of the longitudinal beam 6 is greater than that of the secondary beams 9 and the auxiliary beams 10 to form a differentiated load-bearing skeleton.

[0093] The mobile system includes a power supply component 11, a controller, a warning light 12, and a positioning component 13. The power supply component 11 provides power to the electrical equipment and the multi-dimensional state inversion and early warning module of the transport vehicle 1. The controller (e.g., a PLC or microprocessor), as the core control unit of the transport vehicle's actuators, is electrically connected to the power supply component 11, the warning light 12, the positioning component 13, the multi-dimensional state inversion and early warning module, and the geared motor and hydraulic lifting mechanism of the lifting and rotating system.

[0094] In terms of work logic:

[0095] 1) Regarding the warning function: The controller receives the inversion results output by the algorithm decision layer in real time. When the predicted stress or displacement exceeds the set safety threshold, the controller immediately triggers the warning light 12 to emit an audible and visual alarm signal, and can automatically cut off or reduce the moving power of the mobile system to carry out active safety intervention.

[0096] 2) Regarding the positioning and track-changing functions: The positioning component 13 is used to detect the physical position of the transport vehicle in the track network in real time (e.g., through RFID radio frequency identification or laser positioning and ranging). When the positioning component 13 detects that the transport vehicle has reached the preset track-changing intersection, it feeds back the positioning signal to the controller. Based on this, the controller automatically issues lifting and rotation commands to the lifting and rotating system, thereby realizing precise closed-loop control of heavy-load track-changing actions in a limited space.

[0097] like Figures 4-5 As shown, the units on the adjustable bearing platform 2 are arranged in a stepped stacked manner. Adjacent units on the same layer are spliced ​​together by T-slots to form I-shaped slots, and are laterally locked by built-in L-shaped right-angle connecting seats 14; units on different layers are vertically positioned by positioning pins 15 and fastening bolts 16, so as to achieve adaptive combination according to the size of the tunnel segment.

[0098] like Figure 2 As shown, the lifting and rotating system includes several hydraulic lifting mechanisms 17 installed at the reinforcement node 8. The hydraulic lifting mechanism 17 is connected to the variable cross-section fish belly type main beam 5 through the pressure-bearing guide cylinder 18. The output ends of the hydraulic lifting mechanism 17 on the same side are connected together through the anti-slip head 19. The bottom surface of the anti-slip head 19 is in contact with the external track surface.

[0099] The multidimensional state inversion and early warning module includes:

[0100] The physical sensing layer is used to collect mechanical and attitude data of the transport vehicle in real time during its operation.

[0101] The data-driven layer is used to preprocess the raw data collected by the physical sensing layer and transmit it to the algorithm decision layer.

[0102] The algorithm decision layer embeds a stress and displacement prediction model based on the ConvLSTM architecture. The stress and displacement prediction model is used to invert the structural state of the transfer vehicle in real time during transportation and track changing based on the preprocessed data, and execute structural state judgment logic based on the inversion results to generate early warning signals.

[0103] The visualization and inference layer is used to retrieve the three-dimensional mesh model of the transport vehicle and generate dynamic visualization stress and displacement cloud maps based on the inversion results of the algorithm decision layer.

[0104] The transport vehicle includes a load-bearing frame system, a lifting and rotating system, and a moving system, and is used for transporting and changing the track of the tunnel model;

[0105] The adjustable bearing platform is composed of multiple units with a honeycomb internal structure, and the number of unit combinations and stacking layers are adaptively adjusted according to the size of the tunnel model to be transported.

[0106] The physical sensing layer includes strain gauges, displacement gauges, velocity sensors, load sensors, and position sensors deployed at multiple preset monitoring points on the transport vehicle and adjustable bearing platform.

[0107] The monitoring data collected by the physical sensing layer mainly includes stress and displacement data at key locations of the steel frame 3, the travel speed of the transport vehicle, the lifting speed and jacking force of the hydraulic lifting mechanism 17, the rotation speed of the omnidirectional wheel set, and the total weight of the transport vehicle and the adjustable load-bearing platform; strain gauges are attached to the key stress nodes of the variable cross-section fish-belly type load-bearing main beam 5 and the longitudinal beam 6; displacement gauges are installed at the mid-span of the main beam 5; load sensors are set on the surface of the protective enclosure 4 and the pressure-bearing guide cylinder 18; position sensors and speed sensors are integrated into the moving system;

[0108] like Figure 6 As shown, this embodiment of the invention provides the structural state inversion logic executed by the multidimensional state inversion and early warning module, including the following steps:

[0109] S1, based on the finite element analysis method, simulates the structural state of the transport vehicle under different load conditions and during the track changing process, obtains displacement and stress time-series response data, and constructs a complete structural mechanical performance dataset after preprocessing;

[0110] The specific steps for preprocessing and dataset construction are as follows:

[0111] S101. A finite element model of the transport vehicle was established based on its geometric dimensions, material properties, and connection relationships. After mesh sensitivity analysis, the model was compared and verified with physical test data.

[0112] S102, using experimental design methods, extracts simulation conditions within a range of factor levels covering load magnitude, eccentricity, and adjustable load-bearing platform combination methods;

[0113] S103 uses automated scripts to perform batch solving and data extraction for multiple working conditions. It employs low-pass filtering to remove high-frequency noise, eliminates singular and invalid data, and fills in missing time series data with neighboring values.

[0114] S104 standardizes and maps the processed displacement and stress tensors, integrating them to form a time series dataset of nodal displacement and stress under various working conditions.

[0115] S2, Build a stress and displacement prediction model based on the ConvLSTM architecture, where;

[0116] The weight of the tunnel model, the weight of the adjustable bearing platform, and the bearing position are multimodal input conditions. The spatial topological features and temporal evolution law of the transport vehicle are extracted through neural networks, and the stress and displacement prediction data of each global node are output.

[0117] S3 utilizes a structural mechanics performance dataset to train the stress and displacement prediction model offline, and then verifies and calibrates the trained stress and displacement prediction model based on physical experimental data to ensure that the prediction error range is ≤5%.

[0118] The stress and displacement prediction model based on the ConvLSTM architecture is essentially a dual-branch spatiotemporal feature fusion and physical field decoding network. To achieve the nonlinear mapping from low-dimensional discrete "load location" boundary conditions to high-dimensional continuous "global stress-displacement" physical fields, the specific network architecture and forward inference calculation logic of this stress and displacement prediction model include the following steps:

[0119] S201, Modal input feature tensor quantization;

[0120] Since the state of the transport vehicle is directly driven by dynamic conditions such as the weight of the tunnel model, the weight of the adjustable load-bearing platform, and the load-bearing position, the system first extracts the multimodal input vector at time t. .in, For the real-time weight of the tunnel segments, To provide an adjustable load-bearing platform. The coordinates of the current center of gravity in three-dimensional space.

[0121] To meet the spatial topological information requirements of ConvLSTM, a boundary condition spatial mapping layer is constructed. This layer utilizes a multilayer perceptron and spatial broadcasting mechanism to map low-dimensional vectors... Mapped to a preset space coordinate grid of the transport vehicle, a three-dimensional spatial input tensor containing structural topology information is generated. .in, The number of feature channels after mapping. and This corresponds to the discrete grid dimensions of the horizontal projection plane of the physical structure of the transport vehicle. This input tensor It accurately characterizes the initial distribution of the current load on the variable cross-section fish-belly type main beam and the adjustable load-bearing platform in space.

[0122] S202, Spatiotemporal joint evolution of structural dynamic features based on ConvLSTM;

[0123] The input tensor containing the current load distribution state and the hidden state tensor preserved from the previous time step. With cellular memory tensor It is synchronously fed into the deeply stacked ConvLSTM cascade encoder.

[0124] Unlike traditional fully connected LSTM (Long Short-Term Memory) networks, ConvLSTM introduces convolution operators into its state transition equations. This allows it to extract the dynamic inertial effects during the vehicle's movement while simultaneously capturing the spatial force transmission path of the load within the load-bearing frame system through the local receptive field of the convolution kernel. Internally, it uses multiple gating mechanisms to control the forgetting and updating of physical evolution information. The specific physical computation logic and state transition equations are as follows:

[0125] (1) Gate of Oblivion This is used to assess the impact of the residual strain and structural inertial displacement of the moving vehicle from the previous time step on the current moment. The calculation formula is as follows:

[0126] ;

[0127] (2) Input gate With candidate memory states Used to extract effective local structural response surge features from the current load distribution change:

[0128] ;

[0129] ;

[0130] (3) Cell state renewal By integrating historical effective stress memory with current transient force response, the mechanical evolution state of the global structure of the transport vehicle is updated.

[0131] ;

[0132] (4) Output gate With hidden state Based on the updated cell memory state, a high-dimensional spatiotemporal feature tensor representation of the current time step is output after nonlinear activation.

[0133] ;

[0134] ;

[0135] In the above equation, This represents a two-dimensional spatial convolution operation to extract deformation coupling features between structural nodes; Represents the Hadama product; It is the Sigmoid activation function. It is the hyperbolic tangent activation function; , , and The sequences are the weight kernel matrix and physical bias term of each convolutional layer. These parameters are iteratively optimized and solidified through backpropagation training on the offline finite element simulation dataset (step S3).

[0136] S203, Physical field decoding and inverse normalization reconstruction based on parallel dual-branch network;

[0137] After multiple time steps of iterative iteration, the ConvLSTM encoder outputs a deep hidden state that includes spatiotemporal dynamic features. Since stress and displacement belong to two different physical fields with different dimensions and spatial gradient distribution characteristics, this invention innovatively introduces a parallel dual-branch decoder at the tail end of the stress and displacement prediction model:

[0138] 1. The stress field decoding branch contains two consecutive transposed convolutional layers and a fully connected layer, which will... Sampling is mapped to physical dimensions, and the entire range of the transport vehicle is decoupled and output. Stress prediction vectors of key nodes ;

[0139] 2. The displacement field decoding branch uses independent feature channels for sub-pixel convolution and fully connected layer mapping to output a global displacement prediction vector. .

[0140] To ensure the stability of the aforementioned parallel dual-branch network during actual operation, this invention innovatively introduces a joint smoothing L1 loss function based on multi-task learning during backpropagation optimization. Specifically, since the stress field is prone to numerical singularities at local constraints or abrupt stiffness changes, while the displacement field is relatively smooth, conventional mean square errors impose excessive squared penalties on large errors, easily leading to gradient explosion. Therefore, for the nth node at the target time step t, its predicted stress value... With finite element true value Smoothing loss function between The definition is as follows:

[0141] ;

[0142] Similarly, the smoothing loss function of the displacement branch Defined as:

[0143] ;

[0144] In the formula, and These are preset robustness thresholds for stress and displacement abrupt changes; the advantage of the above physical calculation logic is that when the prediction error is less than... At the threshold, the loss function exhibits a quadratic form (L2 norm), ensuring rapid convergence accuracy of the stress and displacement prediction model under small mechanical fluctuations; when the prediction error is greater than... When the threshold (i.e., under impact or track-changing sudden operating conditions) is reached, the loss function automatically switches to linear (L1 norm), effectively truncating the gradient amplification of abnormal impact signals.

[0145] Furthermore, to balance the dominance of physical fields with different dimensions, stress and displacement, in gradient backpropagation, a multi-task joint loss function including dynamic weighting factors is constructed. :

[0146] ;

[0147] in, and To dynamically adjust hyperparameters for weights; by minimizing the joint loss function. The Adam optimizer is used to update the network weights, so that the stress and displacement prediction model can balance the smoothness of global prediction and the robustness of capturing local singularities in multidimensional state evolution inversion.

[0148] Ultimately, the reasoning generated and The data is transmitted to the multidimensional state inversion and early warning module for dynamic node rendering in the visualization and inference layer.

[0149] like Figure 7 As shown, in a preferred embodiment of the present invention, in order to achieve a high-fidelity mapping from low-dimensional discrete boundary conditions to 39021 finite element nodes across the entire domain, the detailed parameter configurations for each layer of the stress and displacement prediction model are as follows:

[0150] (1) Parameterized implementation of the spatial mapping layer;

[0151] The spatial mapping layer is responsible for transforming the input 1×5 dimensional multimodal vector into a spatial topological tensor. The input L1 layer has 128 channels, and the L2 layer has 16×16×64=16384 output channels. Feature amplification is performed through the fully connected L1 and L2 layers. Subsequently, the one-dimensional feature vector is reconstructed into a three-dimensional spatial tensor of size (16,16,64) using a reshaping operator. This process transforms complex load and position parameters into a gridded feature distribution with physical location attributes.

[0152] (2) ConvLSTM parameters;

[0153] To extract the spatiotemporal evolution features of the transfer vehicle during heavy-load transportation, the encoding region consists of two deeply stacked ConvLSTM layers. ConvLSTM layer 1 uses a 3×3 convolutional kernel with 64 input channels and 64 output channels; ConvLSTM layer 2 uses a 3×3 convolutional kernel with 64 input channels and 128 output channels. Through internal forget gate, input gate, and output gate mechanisms, the local force topology of the structure at the current time step is extracted collaboratively.

[0154] (3) Parallel dual-branch decoder parameters;

[0155] To address the differences in physical field characteristics, independent decoding is performed through parallel branches. In the stress prediction branch, a transposed convolutional layer (4×4 kernel, stride of 2, edge padding size of 1) upsamples the feature map spatial scale from 16×16 to 32×32. Subsequently, a flattening layer transforms the tensor of size (32, 32, 64) into a 65536-dimensional one-dimensional long vector. Finally, a fully connected layer precisely maps the 65536-dimensional features to 39021 discrete nodes, outputting the global stress prediction value.

[0156] The displacement prediction branch first extracts deep deformation features through a feature augmentation convolutional layer (128 input channels, 256 output channels). Then, a sub-pixel convolutional layer (magnification factor of 2, activation function of LeakyReLU) is used to achieve high-fidelity spatial reconstruction, outputting a feature map of size (32, 32, 64). After being transformed by a flattening layer, a fully connected layer outputs the displacement prediction values ​​for all 39021 nodes in the global domain.

[0157] (4) Training monitoring and weight update of the joint loss function;

[0158] To ensure the robustness of multiphysics predictions, a joint smoothing L1 loss function is introduced during the training phase. This function calculates the combined error by comparing the predicted values ​​with the actual finite element values. Figure 6As shown by the connection between "backpropagation and weight update", the generated gradient signal is fed back to the encoder, and the Adam optimizer is used to dynamically iterate the model parameters until the prediction error reaches the preset convergence threshold.

[0159] For example, the scatter regression of stress prediction values ​​and actual values ​​from the stress and displacement prediction model is as follows: Figure 8 As shown, the scatter regression of the predicted and actual displacement values ​​is as follows: Figure 9 As shown in Table 1, the performance of the stress and displacement prediction model on the training and test sets is as follows.

[0160] Table 1:

[0161]

[0162] In a preferred embodiment of the present invention, the visualization and deduction layer constructs a digital twin rendering module that enables bidirectional interaction between the virtual and real worlds. The essence of this digital twin rendering module is to accurately map the discrete, high-dimensional physical quantities output by the algorithm decision layer to the surface of three-dimensional space. To achieve high-fidelity three-dimensional visualization of the mechanical response of the transport vehicle structure, the specific rendering process includes two stages: three-dimensional mesh model reconstruction and data interpolation rendering.

[0163] 1. Analysis and reconstruction of 3D mesh models;

[0164] Due to differences in the underlying data structures between finite element analysis software and 3D graphics rendering engines, mesh topology reconstruction is necessary. The finite element model of the transport vehicle extensively uses four-node shell elements, while the geometric surfaces of the 3D graphics rendering engine are discretized using triangular patches.

[0165] Therefore, the visualization simulation layer first extracts the initial spatial coordinates and element topology connection indexes of all nodes in the finite element model of the transport vehicle using a preset script. Subsequently, based on the triangular facet rendering logic of the rendering engine, the quadrilateral element is divided into two connected triangular facets along the diagonal, and a three-dimensional mesh model with highly consistent topology and geometric dimensions with the original finite element model is generated by following a clockwise vertex arrangement and backface culling mechanism.

[0166] 2. Cloud map rendering based on dynamic normalization and barycenter coordinate interpolation;

[0167] After obtaining a topologically consistent 3D mesh model, the predicted nodal stress and displacement values ​​output by the algorithm's decision layer need to be mapped to the corresponding vertices of the 3D mesh model; the specific rendering calculation steps are as follows:

[0168] (1) Dynamic extreme value normalization: Due to the large amplitude range of the mechanical response data of the transport vehicle under different heavy-load transportation and jacking and rail changing conditions, in order to prevent the color distribution of the rendering, the extreme value normalization process is first performed on the prediction data of all grid vertices at the current time step. The calculation formula is as follows:

[0169] ;

[0170] In the formula, These are the nodal prediction values ​​output by the stress and displacement prediction model; This represents the minimum value of all predicted data under a single working condition. This represents the maximum value of all predicted data under a single working condition. These are the standard values ​​after normalization.

[0171] (2) Centroid Coordinate Interpolation within a Facet: Assigning color values ​​only to discrete vertices results in abrupt color breaks between faces. To achieve a smooth and continuous transition of the mechanical cloud map between adjacent vertices, this invention employs a centroid coordinate interpolation algorithm to continuously color-render the pixels within a triangular facet. Let any vertex of a triangular facet in the 3D mesh be... , , The corresponding normalized predicted values ​​are respectively , , For any point inside the surface... Its physical quantity The interpolation calculation model is defined as follows:

[0172] ;

[0173] Among them, the weighting coefficient , , From point The coordinates of the centroid within the triangle are dynamically determined, satisfying the following:

[0174] ;

[0175] ;

[0176] ;

[0177] (3) Color Gradient Mapping: This calls the programmable graphics pipeline of the rendering engine to construct a mapping from a standard numerical range to a preset RGB color spectrum. The interpolated values ​​are then used to calculate the gradient. Convert it to the corresponding RGB color vector and output it to the screen.

[0178] Through the above-mentioned analytical reconstruction and interpolation rendering mechanism, this system can dynamically display the real mechanical response state of the transport vehicle at the moment of track switching at an extremely high frame rate on a lightweight intelligent control terminal without relying on a large finite element post-processing software.

[0179] S4 deploys the trained stress and displacement prediction model to the multi-dimensional state inversion and early warning module. During operation, it reads the data collected by the physical perception layer in real time and synchronously inverts the structural state of the transport vehicle under the current working conditions through the visualization inference layer.

[0180] The specific implementation logic for creating a visual window is as follows:

[0181] S401, interface framework and functional partition design, uses a dedicated visualization window built on a Python graphical interface framework or other 3D graphics rendering engine. The monitoring interface adopts a modular layout, designed and divided into:

[0182] (1) 3D dynamic rendering interface, embedded with a three-dimensional graphics rendering engine container, used to display the structural mechanical response cloud map of the transport vehicle generated by the visualization inference layer in real time.

[0183] (2) Multi-source data monitoring panel, which refreshes and displays sensor data transmitted from the data-driven layer in real time, including hydraulic jacking force, displacement of key points of the main beam and segment load values.

[0184] (3) Control and interaction area, which integrates camera view switching button, model transparency adjustment slider and warning threshold setting input box.

[0185] S402 provides real-time visualization of multi-dimensional sensor data. Utilizing the preprocessing interface of the data-driven layer, it transforms the acquired physical sensor signals into dynamic line graphs through a graphics rendering component. Simultaneously, a statistical table records the peak stress during the track-changing process in real time, providing a basis for human intervention decisions.

[0186] S403, Inference Logic and Thread Management: To ensure the smoothness of the visualization interface, the system uses a separate computation thread to call the trained ConvLSTM-based stress and displacement prediction model for forward inference. The main thread is responsible for interface rendering, while the inference thread uses real-time acquired input features. The system periodically outputs the predicted values ​​of nodes, enabling simultaneous data collection, inference, and display.

[0187] S404, dynamic mapping and cloud map synchronization, the interface receives the 3D mesh model rendering results output by the visualization inference layer in real time.

[0188] like Figure 10 As shown in the figure, this invention provides an implementation method for an intelligent track-changing transportation system for tunnel models. This embodiment takes the transportation and 90-degree track-changing operation of a certain type of large-span shield tunnel model (weighing 50 tons) as an example. The specific steps are as follows:

[0189] Step 1: System deployment and initialization;

[0190] Before shipment, strain gauges and load sensors were first attached to the key stress points of the variable cross-section fish-belly-shaped load-bearing main beam and longitudinal beams, as well as the pressure-bearing guide cylinder of the transport vehicle, and the position sensors in the moving system were ensured to be in working condition. Subsequently, the monitoring terminal initialized the multi-dimensional state inversion and early warning module, loading the pre-trained stress and displacement prediction model parameters.

[0191] Step 2: Adaptive assembly and load mapping of the adjustable load-bearing platform;

[0192] Based on the geometric dimensions of the tunnel model to be transported, the operators select the appropriate number of honeycomb structural units. Lateral splicing utilizes T-slots on the sides of the units and L-shaped right-angle connectors to achieve lateral locking, ensuring the platform width covers the segment support points; vertical stacking is achieved through multi-layer stacking using positioning pins; the segment transport supports 22 (see [reference]) are installed on the adjustable support platform to hoist the tunnel model. Figure 14 When the boundary conditions are updated, the algorithm decision layer automatically extracts the actual weight Wseg and centroid coordinates (Px, Py, Pz) of the segment fed back by the physical perception layer, and maps the above boundary conditions to the corresponding nodes of the three-dimensional mesh model.

[0193] Step 3, Departure Monitoring and Real-time Inversion:

[0194] After the transport vehicle starts, the physical sensing layer collects sensor data streams at a preset frequency (e.g., 20Hz). The data stream is transmitted from the sensors to the data driving layer for preprocessing, and finally transmitted to the algorithm decision layer. The stress and displacement prediction model, based on the time-series sensor data, simultaneously calculates the real-time stress and displacement of 39,021 nodes across the entire steel frame. The system sets a safety threshold; if the predicted value exceeds the threshold, the terminal displays a red warning and automatically limits the vehicle speed.

[0195] Step four: Rail changing, lifting, and swivel wheel assembly steering;

[0196] When the transport vehicle moves to the preset cross-track area, the positioning component 13 identifies the track-changing node in real time and sends a position arrival signal to the controller. The controller then controls the transport vehicle to stop precisely and automatically triggers the track-changing procedure.

[0197] Lifting process: The controller sends a lifting execution command to the lifting and rotating system, starts the hydraulic lifting mechanism 17, and the hydraulic cylinder extends to drive the anti-slip top 19 to fall onto the surface of the cross track to form a temporary rigid fulcrum; the hydraulic system continues to pressurize, so that the whole vehicle is lifted smoothly until the bottom universal wheel group 7 is completely suspended in the air.

[0198] Steering Execution: Subsequently, the controller sends a steering command to the lifting and rotating system, driving the adaptive steering mechanism of the omnidirectional wheel assembly 7 to activate, making the travel direction of the omnidirectional wheel assembly 7 completely parallel to the direction of the new intersecting track. The steering mechanism adapted to the omnidirectional wheel assembly 7 adopts existing technology; such rotating mechanisms are quite common and will not be described in detail here.

[0199] Step 5: Placement and Secondary Shipment;

[0200] After the lifting and rotating system sends a steering completion signal to the controller, the controller sends a reset command to the lifting and rotating system. The hydraulic lifting mechanism 17 performs controlled pressure reduction, and the anti-slip jack 19 slowly retracts, allowing the caster wheel assembly 7 to smoothly fall onto the cross track. After the anti-slip jack 19 is fully retracted, the controller releases the lock, and the transport vehicle continues to move along the new track direction to the segment installation position. Figures 11-12 As shown.

[0201] Step Six: Full-process digital twin visualization;

[0202] Throughout the entire process described above, the visualization and deduction layer performs the following operations:

[0203] 1. Topology mapping: Maps tens of thousands of node values ​​output by the stress and displacement prediction model to the vertices of the three-dimensional mesh model.

[0204] 2. Rendering: Using color gradient mapping technology, dynamic mechanical cloud maps are generated on the monitoring end.

[0205] 3. Virtual vs. Real: The measured values ​​from the sensors are compared with the predicted values ​​from the model in the sidebar of the interface to ensure that the accuracy of the system inversion can be checked in real time.

[0206] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A tunnel model intelligent transportation and safety status visualization simulation system, characterized in that, include: A multi-dimensional state inversion and early warning module, a transport vehicle, and an adjustable support platform mounted on the transport vehicle; wherein, the multi-dimensional state inversion and early warning module includes: The physical sensing layer is used to collect mechanical and attitude data of the transport vehicle in real time during its operation. The data-driven layer is used to preprocess the raw data collected by the physical sensing layer and transmit it to the algorithm decision layer. The algorithm decision layer embeds a stress and displacement prediction model based on the ConvLSTM architecture. The stress and displacement prediction model is used to invert the structural state of the transfer vehicle in real time during transportation and track changing based on the preprocessed data, and execute structural state judgment logic based on the inversion results to generate early warning signals. The visualization and inference layer is used to retrieve the three-dimensional mesh model of the transport vehicle and generate dynamic visualization stress and displacement cloud maps based on the inversion results of the algorithm decision layer. The transport vehicle includes a load-bearing frame system, a lifting and rotating system, and a moving system, used for transporting and changing the track of the tunnel model; the adjustable load-bearing platform is composed of multiple units with a honeycomb internal structure. The structural state inversion logic executed by the multidimensional state inversion and early warning module includes: S1, Based on the finite element analysis method, the structural mechanical performance dataset of the transport vehicle is obtained through preprocessing and dataset construction; S2, using the structural mechanics performance dataset, train a stress and displacement prediction model based on the ConvLSTM architecture, determine the online inversion calculation logic of the stress and displacement prediction model, and extract the spatial topological features and temporal evolution law of the transport vehicle. S3. The stress and displacement prediction model is trained offline using the structural mechanical performance dataset, and error verification and calibration are performed based on physical test data to ensure that the prediction error range is ≤5%. S4 deploys the trained stress and displacement prediction model to the multi-dimensional state inversion and early warning module, and performs three-dimensional mesh reconstruction and dynamic cloud map rendering through the visualization inference layer; The online inversion calculation logic of the stress and displacement prediction model includes: S201, the processed real-time sensing data, namely the 5-dimensional input vector of the tunnel model weight, the adjustable bearing platform weight, and the current center of gravity coordinates, is reconstructed into a 3D spatial input tensor using a multilayer perceptron and a spatial broadcast mechanism. To accurately characterize the initial spatial distribution of the current load on the transport vehicle; S202, input tensor The hidden state tensor preserved at the previous time step With cellular memory tensor The ConvLSTM is synchronously fed in, and the spatial force transmission path is extracted by introducing a two-dimensional spatial convolution operator.

2. The intelligent transportation and safety status visualization simulation system for tunnel models according to claim 1, characterized in that, The physical sensing layer includes strain gauges, displacement gauges, velocity sensors, load sensors, and position sensors deployed at multiple preset monitoring points on the transport vehicle and adjustable bearing platform. These sensors are used to collect stress and displacement data at the monitoring points of the bearing frame system, the travel speed of the transport vehicle, the lifting speed, lifting force, and rotation speed of the jacking and rotating system, as well as the total weight of the transport vehicle and the adjustable bearing platform. The visualization simulation layer uses node assignment, vertex rendering, and color gradient mapping techniques to map the stress and displacement prediction data output by the algorithm decision layer to the three-dimensional mesh model of the transport vehicle in real time, generating a dynamic visualization cloud map.

3. The intelligent transportation and safety status visualization simulation system for tunnel models according to claim 1, characterized in that, Step S1 further includes: S101. A finite element model is established based on the geometric dimensions, material properties, and connection relationships of the transport vehicle. After performing mesh sensitivity analysis, the model is compared and verified with physical test data. S102 uses experimental design methods to extract multiple simulation conditions within the range of factor level values ​​covering load magnitude, bearing location, and adjustable bearing platform combination mode; S103 uses automated scripts to perform batch solving and data extraction for multiple working conditions. It employs low-pass filtering to remove high-frequency noise, eliminates singular and invalid data, and fills in missing time series data with neighboring values. S104 standardizes and maps the processed displacement and stress tensors, integrating them to form a time series dataset of nodal displacement and stress under various working conditions.

4. The intelligent transportation and safety status visualization simulation system for tunnel models according to claim 1, characterized in that, The stress and displacement prediction model also includes a parallel dual-branch decoding network and a joint loss function; specifically including: S203, Physical field decoding and reconstruction, obtaining the deep hidden state of the ConvLSTM output. The vector is then fed into a parallel dual-branch decoder; the stress branch outputs a global prediction vector through transposed convolution upsampling and flattening mapping; the displacement branch outputs a global prediction vector through subpixel convolution and flattening mapping. S204 introduces a joint smoothing L1 loss function to predict stress values ​​at the target time step. Compared with the true value Its stress smoothing loss function Defined as: when hour, ; when hour, ; Similarly, the displacement branch smoothing loss function Based on threshold Segmented calculation, constructing a system containing dynamic weighting factors and Multi-task joint loss function .

5. The intelligent transportation and safety status visualization simulation system for tunnel models according to claim 1, characterized in that, In step S4, the 3D mesh reconstruction and dynamic rendering of the cloud map include: S401: Extract the node coordinates and element indices of the finite element model pre-established based on the geometric dimensions, material properties and connection relationships of the transport vehicle; divide the quadrilateral elements in the finite element model into triangular patches; and generate a three-dimensional mesh model with the same geometric dimensions as the finite element model. S402, normalize the prediction data of all grid vertices at the current time step; S403 uses a barycentric coordinate interpolation algorithm to perform color rendering on pixels inside the surface. The interpolation calculation model is defined as follows: ； In the formula, , , The normalized predicted values ​​of the face vertices are given, and the weighting coefficients satisfy the following conditions: ; S404 constructs a modular, visual user monitoring interface and integrates trained stress and displacement prediction models. S405 uses the preprocessing interface of the data-driven layer to display the real-time acquired physical sensor data in the form of dynamic line graphs and statistical tables on the interface. S406, invoke the stress and displacement prediction model to perform forward inference, and output the predicted stress and displacement values ​​of the transport vehicle node; S407, in the monitoring interface, synchronously displays a dynamic visualization cloud map generated based on the S403 steps in the form of a 3D cloud map.

6. The intelligent transportation and safety status visualization simulation system for tunnel models according to claim 1, characterized in that, The load-bearing frame system includes two sets of symmetrically arranged variable cross-section fish-belly type load-bearing main beams. The two sets of variable cross-section fish-belly type load-bearing main beams are connected by longitudinal beams and several secondary beams and are equipped with reinforcement nodes. The lifting and rotating system includes several hydraulic lifting mechanisms installed at the reinforcement nodes. The hydraulic lifting mechanisms are connected to the variable cross-section fish-belly type main beam through pressure-bearing guide cylinders. The output ends of the hydraulic lifting mechanisms on the same side are connected together through anti-slip heads. The bottom surface of the anti-slip heads is in contact with the external track surface. The mobile system includes a power supply component, a controller, warning lights, and a positioning component. The power supply component, warning lights, positioning component, lifting and rotating system, and multi-dimensional state inversion and early warning module are electrically or communicatively connected to the controller. The positioning component is used to identify the real-time coordinates and track-changing nodes of the transport vehicle and send the position signal to the controller. The controller is used to receive the early warning signal generated by the algorithm decision layer and the signal from the positioning component, and accordingly control the warning lights to perform audible and visual alarms, and issue track-changing execution commands to the lifting and rotating system. The adjustable bearing platform is a honeycomb stacked structure, comprising several units arranged in a stepped stacked manner. Adjacent units on the same layer are horizontally locked together by T-shaped splicing grooves and L-shaped right-angle connecting seats, while units stacked on different layers are vertically positioned and locked together by positioning components.

7. A method for implementing intelligent transportation of tunnel models, employing the intelligent transportation and safety status visualization simulation system for tunnel models as described in any one of claims 1-6, characterized in that, Includes the following steps: Step 1: Deploy sensors of the physical sensing layer on the transport vehicle and initialize the stress and displacement prediction model parameters of the multidimensional state inversion and early warning module. Step 2: Assemble the adjustable bearing platform according to the size of the tunnel model to be transported, and adjust the assembled adjustable bearing platform to the center of the transport vehicle. Hoist the tunnel model onto the segment transport support installed on the adjustable bearing platform, and simultaneously update the stress boundary conditions of the current working condition to the algorithm decision layer. Step 3: Start the transport vehicle and make it move along the preset trajectory. The algorithm decision layer synchronously inverts the stress and displacement evolution state of the entire domain nodes of the transport vehicle based on the real-time sensor data stream, and issues an early warning when the predicted state exceeds the safety threshold. Step four: After the transport vehicle moves to the rail-changing position, it performs lifting and turning operations through the jacking and rotating system; Step 5: After the turning is completed, the lifting and rotating system performs a reset operation, allowing the transport vehicle to fall onto the cross track and continue moving. Step Six: In the entire transportation and track changing process described above, the sensor measured data and the full-field predicted values ​​output by the stress and displacement prediction model are mapped in real time onto the three-dimensional mesh model of the transport vehicle through the visualization simulation layer, generating a dynamic and visualized stress and displacement cloud map.

Images