Twinning simulation method and system of multi-modal data of shield segment material

Abstract

This invention discloses a twin simulation method and system for multimodal data of tunnel boring machine (TBM) segment materials. The invention relates to the technical field of twin simulation, where multimodal data of TBM segment materials are input into a cross-modal feature decoupling network. Local representation vectors for each mode are extracted and semantically aligned to generate a joint representation tensor for the TBM segment materials, improving the accuracy of this joint representation tensor. Using the joint representation tensor as boundary conditions, the physical constraint neural network built into the digital twin performs a forward solution, generating a corresponding multiphysics response cloud map. Key damage features are extracted from the multiphysics response cloud map, and combined with measured data of the TBM segment materials to determine the simulation deviation content. This triggers online updates of the cross-modal feature decoupling network and the physical constraint neural network, improving the accuracy of the actual physical state of the TBM segment materials.

Description

Technical Field

[0001] This invention relates to the technical field of twin simulation, and in particular to a twin simulation method and system for multimodal data of tunnel segment materials. Background Technology

[0002] As the core prefabricated component of underground tunnel engineering, the mechanical integrity of shield tunnel segments directly determines the safe service life of the tunnel structure. Under extreme conditions such as complex geology and high water pressure, there is a complex cross-scale coupling relationship between the evolution of microscopic defects inside the segments and macroscopic mechanical failure. In order to explore the damage mechanism of segment materials, digital twin simulation technology has been gradually introduced into the full life cycle assessment of segment structures.

[0003] In existing technologies, traditional digital twin simulation methods for tunnel boring machine (TBM) segments typically rely on single-modal data-driven approaches or purely finite element numerical simulations. For example, conventional simulations only use the macroscopic mechanical response parameters of the material or simple geometric defects as input boundary conditions for the simulation model. However, TBM segments are a typical heterogeneous multiphase composite material, which undergoes complex internal topological variations and dynamic microcrack propagation during the preparation and loading stages. Single-modal data cannot fully represent the deep physical connections between the three elements of "spatial geometric defects - dynamic damage mechanism - macroscopic mechanical response" within the material, affecting the accuracy of the joint characterization tensor of the TBM segment material and resulting in low accuracy of the true physical state of the TBM segment material. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a twin simulation method and system for multimodal data of shield tunnel segment materials.

[0005] This invention provides a twin simulation method for multimodal data of tunnel segment materials, including:

[0006] X-ray CT voxel data of shield tunnel segment materials during the preparation stage were obtained, and spatiotemporal alignment was performed by combining the acoustic emission waveform data of shield tunnel segments and the mechanical response data of shield tunnel segment materials to determine the multimodal data of shield tunnel segment materials.

[0007] Multimodal data is input into a cross-modal feature decoupling network, and local representation vectors of each modality are extracted. The local representation vectors are then semantically aligned to generate a joint representation tensor of the shield tunnel segment material.

[0008] The joint representation tensor is input into a preset digital twin, and the twin simulation of the joint representation tensor is triggered. The joint representation tensor is used as the boundary condition, and the physical constraint neural network built into the digital twin performs a forward solution to generate the corresponding multiphysics response cloud map.

[0009] Key damage features are extracted from the multiphysics response cloud map, and the simulation deviation is determined by combining the measured data of the shield tunnel segment material. The online update of the cross-modal feature decoupling network and the physical constraint neural network is triggered, so that the output of the digital twin converges to the real physical state of the shield tunnel segment material.

[0010] This invention provides a twin simulation system for multimodal data of tunnel boring machine (TBM) segment materials. This twin simulation system is applied to the aforementioned twin simulation method for multimodal data of TBM segment materials. The twin simulation system for multimodal data of TBM segment materials includes:

[0011] The multimodal data module is used to acquire X-ray CT voxel data of shield tunnel segment materials during the preparation stage, and combine it with the acoustic emission waveform data of shield tunnel segments and the mechanical response data of shield tunnel segment materials for spatiotemporal alignment, thereby determining the multimodal data of shield tunnel segment materials;

[0012] The joint representation tensor module is used to input multimodal data into the cross-modal feature decoupling network, extract the local representation vectors of each mode, and perform semantic alignment on the local representation vectors, thereby fusing them to generate a joint representation tensor of the shield tunnel segment material.

[0013] The twin simulation module is used to input the joint representation tensor into a preset digital twin and trigger the twin simulation of the joint representation tensor. Using the joint representation tensor as boundary conditions, the physical constraint neural network built into the digital twin performs forward solving to generate the corresponding multiphysics response cloud map.

[0014] The online update module is used to extract key damage features from the multiphysics response cloud map, combine them with the measured data of the shield tunnel segment material to determine the simulation deviation, and trigger the online update of the cross-modal feature decoupling network and the physical constraint neural network, so that the output of the digital twin converges to the real physical state of the shield tunnel segment material.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] (1) Obtain X-ray CT voxel data of shield tunnel segment material during the preparation stage, and combine the acoustic emission waveform data of shield tunnel segment and the mechanical response data of shield tunnel segment material for spatiotemporal alignment to determine the multimodal data of shield tunnel segment material; input the multimodal data into the cross-modal feature decoupling network, extract the local representation vector of each mode respectively, and perform semantic alignment on the local representation vector to fuse and generate the joint representation tensor of shield tunnel segment material. The multimodal data of shield tunnel segment material is introduced, and the local representation vector of each mode is further controlled, which improves the accuracy of the joint representation tensor of shield tunnel segment material.

[0017] (2) Input the joint representation tensor into the preset digital twin and trigger the twin simulation of the joint representation tensor. Using the joint representation tensor as the boundary condition, the physical constraint neural network built into the digital twin performs forward solving to generate the corresponding multiphysics response cloud map. Extract the key damage features in the multiphysics response cloud map, combine them with the measured data of the shield tunnel segment material to determine the simulation deviation content, and trigger the online update of the cross-modal feature decoupling network and the physical constraint neural network. This makes the output result of the digital twin converge to the real physical state of the shield tunnel segment material, realize the twin simulation of the joint representation tensor, fully consider the simulation deviation content, the cross-modal feature decoupling network and the physical constraint neural network, and improve the accuracy of the real physical state of the shield tunnel segment material. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the twin simulation method for multimodal data of tunnel segment materials in an embodiment of the present invention;

[0019] Figure 2 This is a flowchart illustrating step S11 in the twin simulation method for multimodal data of tunnel segment materials in this embodiment of the invention.

[0020] Figure 3 This is a flowchart illustrating step S12 in the twin simulation method for multimodal data of tunnel segment materials in this embodiment of the invention.

[0021] Figure 4 This is a flowchart illustrating step S13 in the twin simulation method for multimodal data of tunnel segment materials in this embodiment of the invention.

[0022] Figure 5 This is a flowchart illustrating step S14 in the twin simulation method for multimodal data of tunnel segment materials in this embodiment of the invention.

[0023] Figure 6 This is a schematic diagram of the structural composition of the twin simulation system for multimodal data of tunnel segment materials in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] Please see Figures 1 to 6 A twin simulation method for multimodal data of tunnel boring machine (TBM) segment materials is proposed and applied to twin simulation scenarios. The twin simulation method for multimodal data of TBM segment materials includes:

[0026] Step S11: Obtain X-ray CT voxel data of shield tunnel segment material during the preparation stage, and combine it with the acoustic emission waveform data of shield tunnel segment and the mechanical response data of shield tunnel segment material for spatiotemporal alignment, thereby determining the multimodal data of shield tunnel segment material;

[0027] Step S12: Input the multimodal data into the cross-modal feature decoupling network, extract the local representation vectors of each modality, and perform semantic alignment on the local representation vectors to fuse and generate the joint representation tensor of the shield tunnel segment material;

[0028] Step S13: Input the joint representation tensor into the preset digital twin and trigger the twin simulation of the joint representation tensor. Using the joint representation tensor as the boundary condition, the physical constraint neural network built into the digital twin performs forward solving to generate the corresponding multiphysics response cloud map.

[0029] Step S14: Extract key damage features from the multiphysics response cloud map, combine them with the measured data of the shield tunnel segment material to determine the simulation deviation content, and trigger the online update of the cross-modal feature decoupling network and the physical constraint neural network so that the output of the digital twin converges to the real physical state of the shield tunnel segment material.

[0030] refer to Figure 2 In step S11, the specific steps are as follows:

[0031] S111: Acquire X-ray CT voxel data, acoustic emission waveform data, and mechanical response data of shield tunnel segment materials during the preparation stage; perform noise reduction and reconstruction on the X-ray CT voxel data to extract the corresponding topological features; perform wavelet packet decomposition on the acoustic emission waveform data to extract frequency band energy features; and simultaneously identify the mechanical response data to extract the corresponding stress-strain data.

[0032] S112: The mechanical response data is synchronized with the X-ray CT voxel data and acoustic emission waveform data by time axis interpolation. The microscopic topological structure features, frequency band energy features and stress-strain data are jointly registered to eliminate the spatiotemporal scale differences and phase delays between multimodal data and generate multimodal data of shield tunnel segment materials. This multimodal data shows spatial coordinate alignment, time series synchronization and physical state correlation.

[0033] In the embodiments of this application, X-ray CT voxel data, acoustic emission waveform data, and mechanical response data of the shield tunnel segment material during the preparation stage are acquired; the X-ray CT voxel data are denoised and reconstructed to extract the corresponding topological features; the acoustic emission waveform data are decomposed by wavelet packet to extract the frequency band energy features; at the same time, the mechanical response data are identified to extract the corresponding stress-strain data, thus introducing the corresponding stress-strain data.

[0034] At this point, the system acquires high-resolution X-ray CT voxel data of the tube segment sample during preparation and loading. Given that industrial CT scanning is susceptible to X-ray scattering and quantum noise interference, and that a large number of artifacts exist in the original voxel matrix, the system then uses a three-dimensional block-matched filtering method based on nonlocal means to denoise the voxel data, so as to suppress high-frequency noise while preserving the edge sharpness of the transition zone between aggregate and mortar.

[0035] For noise reduction processing, when reconstructing X-ray CT voxel data, the system employs a three-dimensional block matching filter method based on non-local means. This method divides the three-dimensional voxel matrix to be processed into several three-dimensional cube image blocks with a side length of 8 voxels. The search window has a side length of 32 voxels, and other image blocks similar to the current image block are searched within this search window. For each similar block, the system further calculates the corresponding similarity weight based on its Euclidean distance to the current block. The system then performs a weighted average of the gray values ​​of all similar blocks within the search window and replaces the original gray value of the center voxel of the current image block with this weighted average. To balance noise reduction efficiency and edge preservation, the system can iteratively execute the above filtering process twice, with the image block side length adjusted to 16 voxels in the second iteration. Through the above parameter configuration, high-frequency artifacts caused by X-ray scattering and quantum noise are effectively suppressed while preserving the edge sharpness of the aggregate-mortar interface transition zone.

[0036] After noise reduction, the system performs multiphase segmentation on the reconstructed 3D voxel matrix based on a preset grayscale gradient threshold, accurately distinguishing the coarse aggregate phase, mortar matrix phase, and initial microporous phase. Based on this, the system extracts the corresponding topological features, specifically representing the connected domain volume, spatial distribution density matrix, and pore morphology orientation tensor of each phase in 3D space, thereby transforming discrete voxel grayscale values ​​into a set of pure topological parameters characterizing the non-uniform solid geometry within the material. The system acquires high-frequency acoustic emission waveform data from an array of acoustic emission sensors on the surface of the tunnel lining segments. Considering the complex damage mechanisms and highly overlapping waveforms in the concrete material of the tunnel lining segments, the system employs wavelet packet decomposition technology to perform frequency domain analysis on the aforementioned acoustic emission waveform data.

[0037] At this point, the system selects the Daubechies wavelet basis function, which has compact support and orthogonality, to perform three to five layers of wavelet packet complete binary tree decomposition on the acoustic emission waveform signal, dividing the full-band signal into multiple orthogonal sub-bands without redundancy or omission. The system calculates the square integral value of the reconstructed signal in each sub-band layer by layer to obtain the signal energy corresponding to each frequency band. Based on this, the system extracts the frequency band energy characteristics, which are specifically represented as the vector distribution of the proportion of energy in different frequency bands to the total energy. The proportion of low-frequency band energy is characterized by the tensile propagation of microcracks in the mortar matrix inside the concrete, while the proportion of high-frequency band energy is characterized by the shear damage caused by coarse aggregate fracture or interface friction, thereby realizing the mapping of acoustic emission signal to quantitative damage mechanism characteristics.

[0038] The system synchronously acquires the mechanical response data of the shield tunnel segment material fed back by the servo hydraulic testing machine. This raw data includes load and displacement values ​​that change continuously over time. The system performs system-level error identification and compensation on the above mechanical response data, eliminating the initial displacement zero-point drift caused by the slippage of the testing machine fixture, and uses Butterworth low-pass filtering to filter out high-frequency burrs of mechanical vibration during the loading process. After data cleaning, the system converts the load-displacement time series data into engineering stress-engineering strain time series data based on the actual gauge length of the segment specimen, and further converts it into true stress-logarithmic plastic strain data based on the basic equations of plasticity. Finally, the system extracts the corresponding stress-strain data, which serves as the unique external absolute mechanical boundary anchor point in the multimodal dataset, accurately calibrating the macroscopic mechanical response trajectory of the material from elastic deformation, yield hardening to softening failure after peak value.

[0039] Specifically, the system executes step S111 on a scaled-down C50 reinforced concrete segment specimen under extreme soil and water pressure simulation. Through an X-ray CT sub-step, the system sensitively captures the beaded, millimeter-sized pore groups inside the specimen before loading, caused by insufficient compaction during pouring and vibration. These are then transformed into a three-dimensional topological feature matrix characterizing the stress concentration source. As external loads are applied progressively, when invisible microcracks begin to emerge inside the segment, the system uses an acoustic emission wavelet packet decomposition sub-step to analyze the transient waveform captured by the sensor in real time. It extracts the frequency band energy feature vector showing a sudden increase in energy rate within the 80kHz-120kHz band, which is directly identified by the system as tensile peeling damage at the aggregate-mortar interface. Simultaneously, the mechanical response data identification sub-step outputs real-time stress-strain data showing a stress reaching 45MPa and strain beginning to exhibit nonlinear accelerated growth.

[0040] The three types of features extracted above are rigorously input into a cross-modal feature decoupling network. The digital twin is no longer a blind calculation based on ideal homogeneous materials, but directly maps the above "beaded pore topology features" into the initial boundary conditions of local stiffness reduction in the physical solution mesh, uses the "80kHz-120kHz frequency band energy features" as the dynamic evolution constraint of microcrack propagation direction, and uses "45MPa real stress-strain data" as the macroscopic load scale. Under the high-fidelity joint drive of this multimodal data, the physical constraint neural network in the digital twin accurately reproduces the entire process of tunnel segment crushing: "tensile crack initiation at micropores > crack propagation around the aggregate interface > final formation of macroscopic penetrating splitting failure surface". The generated multiphysics response cloud map has an error of less than 3% with the final failure mode of the real physical specimen, thus realizing a high-confidence twin mapping and early warning of the failure mechanism of shield tunnel segments under complex working conditions.

[0041] Furthermore, the mechanical response data is synchronized with the X-ray CT voxel data and acoustic emission waveform data through time axis interpolation. The microscopic topological features, frequency band energy features and stress-strain data are jointly registered to eliminate the spatiotemporal scale differences and phase delays between multimodal data, generating multimodal data of the shield tunnel segment material. This multimodal data presents spatial coordinate alignment, time series synchronization and physical state correlation.

[0042] At this point, after extracting the underlying data of each modality, the system performs time axis interpolation synchronization to address the severe time scale mismatch between the mechanical response data and the X-ray CT voxel data and acoustic emission waveform data. Since the sampling frequency of the mechanical response data is usually in the range of Hertz to kilohertz and is a continuous time series, while a single three-dimensional scan of the X-ray CT voxel data takes several minutes to tens of minutes to form extremely low-frequency discrete time frames, the system uses the absolute continuous time axis of the mechanical test as the main time reference and employs cubic spline interpolation to densify and extend the time dimension of the discrete CT voxel data frames in order to construct a continuously changing virtual CT state sequence.

[0043] Optionally, the system first identifies four characteristic time points where the material stiffness changes significantly through a set of continuous mechanical response data within the same loading cycle: the linear elastic segment termination point, the crack initiation initiation point, the midpoint of stable crack propagation, and the peak stress point. The system uses the above four characteristic time points as control nodes for cubic spline interpolation, and uses the discrete CT scan frame that is closest to these time points as the reference frame.

[0044] Between two adjacent CT reference frames, the system assumes that the evolution of the microstructure follows a stiffness degradation path determined by the mechanical response data. Specifically, in the interpolated virtual CT state sequence, the topological parameters of each phase are linearly weighted and fused according to the modulus decay ratio of the corresponding time period in the mechanical response data. Taking two actual CT scan frames at the 10-minute and 18-minute intervals as examples, if the mechanical response data shows that the secant modulus of the material has decreased from its initial value to 80% of its initial value at the 12-minute interval, then in the virtual CT state at the 12-minute interval, the system assigns a 20% weight to the microstructure features of the 10-minute frame and an 80% weight to the microstructure features of the 18-minute frame, and then weights and sums the gray values ​​of the corresponding voxel positions in the two frames according to this weighting coefficient. The physical basis of this method is that, under quasi-static loading conditions, the evolution rate of the material's microstructure is monotonically correlated with the rate of change of the macroscopic mechanical response; that is, the faster the modulus decreases, the faster the rate of accumulation of micro-damage.

[0045] The system ultimately applies the aforementioned weighted combination only to the time intervals between two adjacent actual CT scan frames. The interpolated virtual CT sequence is used solely to drive the intermediate state deduction of the physical constraint neural network within the digital twin, and is not output as independent external verification data. To verify the validity of the interpolation results, the system can also compare the virtual CT frame at the interpolation node with the actual low-resolution fast CT pre-scan frames captured within two minutes before and after that node. If the average grayscale deviation between the two is less than five percent, the interpolation result is deemed to have physical credibility.

[0046] Meanwhile, to address the fixed phase delay generated by acoustic emission waveform data in wavefront propagation, sensor coupling, and signal conditioning circuits, the system uses the precise moment of stress surge in the mechanical response data as the trigger zero point, calculates the time difference matrix of the acoustic emission signal arriving at each sensor in reverse, and subtracts the wavefront propagation time in reverse through ray tracing, thereby strictly translating and aligning the starting phase of the acoustic emission characteristic sequence to the same absolute time zero point of mechanical loading, achieving millisecond-level strict synchronization of the three-modal data in the time series.

[0047] After eliminating time scale differences and phase delays, the system then performs three-dimensional joint registration of microscopic topological features, frequency band energy features, and stress-strain data. The system pre-embeds or presents high-density annular metal reference marks on the surface of the specimen, identifies the geometric center of the mark in the CT voxel data through sub-pixel edge extraction, and uses it as the origin and axial reference of the global cuboid physical coordinate system.

[0048] For frequency band energy characteristics and stress-strain data lacking native spatial coordinates, the system performs dimensionality reduction mapping: the stress-strain data is regarded as a global macroscopic scalar field bound to the origin of the physical coordinate system; for acoustic emission radio frequency band energy characteristics, the system introduces a Geiger joint positioning method based on a multi-sensor array, and calculates the three-dimensional spatial coordinates of the acoustic emission source of the frequency band energy characteristics inside the tube segment based on the waveform time difference received by each sensor and the preset material wave velocity model; the system uniformly resamples the above acoustic emission source spatial coordinate matrix and global mechanical scalar and rigidly transforms it to a physical coordinate system with the CT reference mark as the origin, thereby eliminating the spatial misalignment of the three-mode data caused by sensor position offset or specimen clamping tilt, and achieving a perfect alignment of spatial coordinates.

[0049] To address the lack of native spatial coordinates for the acoustic emission frequency band energy characteristics, the system introduces a Geiger joint localization method based on a multi-sensor array. This method inversely calculates the three-dimensional spatial coordinates of the acoustic emission source within the tunnel segment for that frequency band energy characteristic. Specifically, the system pre-positions six acoustic emission sensors on the surface of the tunnel segment specimen. Their spatial coordinates are as follows: Sensor 1 (0 mm, 0 mm, 0 mm), Sensor 2 (100 mm, 0 mm, 0 mm), Sensor 3 (200 mm, 0 mm, 0 mm), Sensor 4 (0 mm, 100 mm, 0 mm), Sensor 5 (100 mm, 100 mm, 0 mm), and Sensor 6 (200 mm, 100 mm, 0 mm). All these coordinates are established in a right-handed Cartesian coordinate system with the lower left front corner of the tunnel segment as the origin.

[0050] For the material wave velocity model, considering that the shield tunnel segments in this embodiment are made of C50 concrete, the longitudinal wave velocity is 4200 m / s under stress-free conditions, and rises to 4450 m / s under 50 MPa high water pressure confining pressure due to pore compression. Based on this, the system establishes a linear relationship between wave velocity and stress state, that is, the current wave velocity is equal to the reference wave velocity of 4200 m / s plus the current stress value multiplied by 5 m / MPa.

[0051] Regarding initial value settings, the system uses the coordinates of the sensor that receives the acoustic emission signal earliest among the six sensors as the geometric center of the initial source coordinates. The initial earthquake time is set as the earliest arrival time minus the time difference between the sensor and the initial coordinates, divided by the current wave velocity. During the positioning iteration process, the system establishes six hyperbolic equations, each representing that the arrival time difference between the source and a certain sensor and the source and the reference sensor is equal to the difference between the two distances divided by the current wave velocity. The system uses the least squares method to solve this overdetermined system of equations. After each iteration, the correction amount of the source coordinates is calculated. When the change in the Euclidean distance of the source coordinates between two consecutive iterations is less than 0.5 mm, convergence is determined, and the iteration stops. If convergence is not achieved after more than fifty iterations, the system abandons the positioning event to avoid wasting computational resources.

[0052] After completing the spatiotemporal alignment, the system further establishes a physical state causal mapping between modes to generate the final multimodal data. The system uses "stress-strain data" with clear mechanical and physical meaning as state anchors to strongly bind "microscopic topological features" and "frequency band energy features" at the same timestamp. At this point, at any discrete synchronization node in the time dimension, the system extracts the current stress state value at that moment and extracts the interpolated CT topological matrix corresponding to that moment and the frequency band energy feature point cloud of all acoustic emission sources falling within that time window in the spatial dimension. The system concatenates and encapsulates the above three feature vectors that are at the same absolute moment, in the same physical coordinate system, and have mechanical causal relationships, and finally generates a shield tunnel segment material multimodal data tensor that presents spatial coordinate alignment, time series synchronization, and physical state correlation, completely eliminating the semantic island effect between heterogeneous data.

[0053] Specifically, the tunnel segment specimen is being tested in a servo-loaded testing machine simulating high water pressure. When the system performs time axis interpolation synchronization, facing the high-frequency stress-strain data generated by the testing machine continuously loading at a speed of 0.5 mm per second, and the extremely low-frequency discrete CT images caused by the CT scanner needing to stop for 5 minutes after scanning each tomography, the system smoothly extrapolates the static CT scan frame at the 10th minute to the 11th minute on the time axis through cubic spline interpolation, constructing a continuous dynamic evolution sequence of the topological defects inside the tunnel segment during this period. At the same time, for the extremely weak acoustic emission signal generated by the closure of microcracks inside the concrete under deep water pressure, the system accurately subtracts the phase delay of the sound wave propagating in water-saturated concrete, so that the moment of sudden increase in acoustic emission radio frequency band energy and the moment of the appearance of a small nonlinear compaction segment on the stress-strain curve are aligned at the microsecond level.

[0054] The system uses metal positioning rings on the edge of the specimen to precisely anchor the local water-rich pore group detected by CT scan at the position (150mm, 80mm, 200mm) in the physical coordinate system. At this time, the group of 80kHz-120kHz high-frequency acoustic emission sources, which represent the tensile peeling of microcracks at the pore edge, calculated by Geiger joint positioning method, was originally based on the local coordinate system of the sensor array. The system forcibly translated and mapped it to the above-mentioned global physical coordinate system. It was found that these acoustic emission sources surrounded the CT pore group at (150mm, 80mm, 200mm) with extremely high density in space.

[0055] The system extracts the state at a specific moment of 11 minutes and 15 seconds: taking the "stress of 50MPa and strain of 0.0025" at this moment as the physical anchor point, and binding it with the "microscopic pore topology deformation matrix centered at (150mm, 80mm, 200mm)" interpolated at this moment, as well as the "80kHz-120kHz frequency band energy characteristic cloud" that bursts in the same spatial coordinate neighborhood at this instant; the multimodal data generated thus clearly proclaims an irrefutable physical fact to the subsequent digital twin: at an absolute time of 11 minutes and 15 seconds, at a specific coordinate inside the segment, due to the high water pressure load of 50MPa, the originally static microscopic pore topology begins to undergo real physical degradation through the high-frequency tensile crack mechanism. This high-fidelity multimodal data is directly used as the input of the physical constraint neural network of the digital twin, driving it to accurately reproduce the entire process of the segment's evolution from the initiation of local microscopic defects to macroscopic crushing under water pressure.

[0056] refer to Figure 3 In step S12, the specific steps are as follows:

[0057] S121: Multimodal data is input into a cross-modal feature decoupling network. The cross-modal feature decoupling network includes a 3D convolutional twin branch, a temporal graph convolutional branch, and a fully connected feature mapping branch that process X-ray CT voxel data, acoustic emission waveform data, and mechanical response data, respectively. The corresponding local representation vectors are extracted through each branch, and a modal orthogonal projection module is introduced to perform semantic alignment on each local representation vector in order to output the corresponding common semantic features.

[0058] S122: Using the local representation vectors corresponding to the mechanical response data as semantic anchors, the semantic alignment loss is determined through the attention mechanism of cross-modal contrastive learning, so as to minimize the distance between the local representation vectors of different modalities representing the same physical damage state in the latent feature space. The common semantic features are spliced ​​and fused, and the temporal context is aggregated through a gated recurrent unit to generate the joint representation tensor of the shield tunnel segment material.

[0059] In the embodiments of this application, multimodal data is input into a cross-modal feature decoupling network. The cross-modal feature decoupling network includes a three-dimensional convolutional twin branch, a temporal graph convolutional branch, and a fully connected feature mapping branch that respectively process X-ray CT voxel data, acoustic emission waveform data, and mechanical response data. The corresponding local representation vectors are extracted through each branch, and a modal orthogonal projection module is introduced to perform semantic alignment on each local representation vector to output the corresponding common semantic features.

[0060] At this point, the system calls the three-dimensional convolutional twin branch to process X-ray CT voxel data with complex spatial structures. This branch adopts a residual network architecture containing three-dimensional convolutional kernels and batch normalization layers. By setting the stride and padding strategy, it performs spatial downsampling and feature channel upsampling operations on the input three-dimensional microscopic topological feature matrix.

[0061] In this process, the twin branches utilize a two-stream structure with shared weights to compare and process the voxel data of the current load step and the initial state, respectively, to filter out the static background skeleton and highlight the dynamically evolving microcracks and pore topological distortions. After multi-level 3D feature extraction and global average pooling operations, the system compresses the high-dimensional spatial feature map into a fixed-dimensional latent space, thereby outputting a 3D topological local representation vector that can accurately characterize the distribution of non-uniform geometric defects inside the tube segment. The system inputs the frequency band energy features extracted and spatiotemporally registered in the previous stage into the time-series graph convolution branch. Given the spatial discreteness and temporal suddenness of acoustic emission events, the system constructs a dynamic topological graph structure from the spatial coordinates of multiple acoustic emission sources within a specific time window. The node features are defined as the frequency band energy proportion vector of each acoustic emission source, and the edge features are defined as the spatial Euclidean distance and temporal causal relationship between the acoustic emission sources.

[0062] The system employs a temporal graph convolutional network with a self-attention mechanism to aggregate and update neighbor node information on the topology graph in order to capture the co-evolution law of adjacent acoustic emission events in spatial clustering and energy conduction during the microcrack propagation process. The system performs dimensionality reduction and integration of the node features of the entire graph through a graph-level readout function, and outputs a local acoustic emission representation vector that can reflect the spatiotemporal propagation state of the dynamic damage mechanism inside the material.

[0063] The system flattens the registered stress-strain data sequence within a continuous time window and inputs it into a fully connected feature map branch. This branch consists of multiple hidden layers with nonlinear activation functions. Through layer-by-layer linear transformation and nonlinear mapping, it mines deep macroscopic mechanical indices such as loading rate, stiffness degradation rate, and energy dissipation integral hidden in the stress-strain sequence. The fully connected branch projects one-dimensional time-series mechanical parameters into a latent feature space with the same dimensions as the two branches mentioned above, thereby outputting a mechanical local characterization vector that can macroscopically calibrate the current loading stage and overall mechanical response level of the material.

[0064] After acquiring three heterogeneous local representation vectors, the system immediately introduces a modal orthogonal projection module to perform a strict semantic alignment operation to eliminate the feature space distribution offset caused by the different physical nature of the data sources. This module initializes a set of mutually orthogonal common semantic basis vectors in the latent space and uses a loss function based on cross-modal contrastive learning to forcibly increase the distance between feature classes under different physical states. The system projects the 3D topological local representation vectors, acoustic emission local representation vectors, and mechanical local representation vectors onto this set of orthogonal basis vectors respectively, stripping away the redundant environmental noise and individual-specific information unique to each mode, such as CT ray hardening artifacts and acoustic emission mechanical friction noise, retaining only the common semantic components that can all point to the same physical state of material damage. After orthogonal constraints and feature reconstruction, the system finally outputs common semantic features that eliminate the modal gap and are highly aligned, providing a joint driving source with physical consistency for subsequent multiphysics twin solutions.

[0065] Specifically, when the segment specimen is subjected to simulated extreme high water pressure, such as the 50MPa stress point on the stress-strain curve, multimodal data floods into the cross-modal feature decoupling network; the three-dimensional convolutional twin branch receives CT voxel data from the current time and the initial time. Through the twin comparison mechanism, this branch keenly removes a large area of ​​intact aggregate background and accurately extracts a high-dimensional vector. This vector specifically presents the pure three-dimensional spatial topological distortion state of "micropore groups being compacted and bifurcating microcracks emerging" in a certain area deep on the tensile side of the segment.

[0066] Meanwhile, the temporal graph convolution branch processes acoustic emission data at the same time. Faced with dense and chaotic acoustic emission signals under high water pressure, this branch constructs a dynamic topological graph of discrete acoustic emission sources. Through graph convolution aggregation, a high-dimensional vector is extracted. This vector accurately represents the spatiotemporal propagation mechanism of acoustic emission events in the same region undergoing spatial clustering in a very short time and the explosive series growth of the proportion of high-frequency tensile crack energy in the 80kHz-120kHz range.

[0067] In addition, the fully connected eigenmap branch maps the seemingly flat stress-strain curve into a high-dimensional vector, which implicitly contains the special mechanical state caused by the water pressure confining pressure effect, where "the macroscopic apparent stiffness has not decreased significantly, but the internal plastic strain energy has begun to accumulate locally". At this point, without semantic alignment, these three vectors would be far apart in the latent space, and the digital twin would not be able to understand that they describe the same physical event. The modal orthogonal projection module then starts, projecting the "microcrack topological distortion vector", "high-frequency energy clustering vector" and "local plastic accumulation vector" onto a set of orthogonal bases simultaneously. After orthogonal constraint and contrastive learning alignment, the module forcibly removes the grayscale noise of CT and the machine vibration noise of acoustic emission, so that the heterogeneous information originally described by each other is forcibly "translated" into the same language - that is, outputting a highly condensed common semantic feature. The only semantic meaning that the common semantic feature conveys to the digital twin is: "Under the current high water pressure of 50MPa, in the specific coordinate domain of the tension side of the segment, although the macroscopic mechanics has not yet become unstable, the microscopic topological defects have been forced to activate and have begun to undergo substantial tensile damage clustering."

[0068] Furthermore, using the local representation vectors corresponding to the mechanical response data as semantic anchors, the semantic alignment loss is determined through the attention mechanism of cross-modal contrastive learning to minimize the distance between the local representation vectors of different modalities representing the same physical damage state in the latent feature space. Common semantic features are spliced ​​and fused, and then time-series context aggregation is performed through a gated recurrent unit to generate a joint representation tensor of the shield tunnel segment material. Multimodal data of the shield tunnel segment material is introduced, and the local representation vectors of each modality are further controlled, thereby improving the accuracy of the joint representation tensor of the shield tunnel segment material.

[0069] At this point, after acquiring the local representation vectors of each modality, the system immediately constructs a semantic alignment constraint mechanism based on mechanical priors. The system explicitly sets the local representation vectors corresponding to the mechanical response data as non-drifting semantic anchors, using these anchors as the benchmark reference system for the absolute physical state. When calculating the semantic alignment loss, the system introduces an attention mechanism for cross-modal contrastive learning, using the mechanical local representation vector as the query vector and the CT topological local representation vector and the acoustic emission local representation vector as key-value pairs. By calculating the scaling dot product attention score between the query vector and the key-value pairs, the system dynamically evaluates the semantic contribution of different modal features to the current macroscopic mechanical state. The system aims to minimize the Euclidean or cosine distance between the local representation vectors of each modality at the same timestamp in the latent feature space. It constructs a contrastive loss function in combination with attention weights, forcibly driving the CT features and acoustic emission features representing the same physical damage state to converge and move towards the mechanical anchor in the latent space on a high-dimensional manifold, thereby eliminating the semantic residuals between heterogeneous modalities.

[0070] The system performs feature-level splicing and fusion on the common semantic features of each channel that have eliminated the modal gap. The system will perform a head-to-tail concatenation operation on the channel dimension on the common semantic features of CT topology, common semantic features of acoustic emission radio frequency band energy, and common semantic features of mechanical response that are dimensionally consistent and semantically aligned in the latent feature space. This splicing process will reintegrate the spatial geometric distortion information, dynamic damage mechanism information, and macroscopic mechanical boundary information that were originally scattered in different modes into a high-dimensional multi-feature matrix. In order to prevent feature degradation caused by dimensional redundancy and linear correlation in the spliced ​​feature matrix, the system will immediately follow the splicing layer with a nonlinear projection network containing a bottleneck structure to perform dimensionality reduction and nonlinear activation mapping on the high-dimensional splicing matrix and extract a highly condensed multimodal fusion feature vector.

[0071] Given that the damage evolution of tunnel segment materials is a dynamic temporal process with strong historical dependence, the system ultimately inputs the aforementioned multimodal fusion feature vectors into a gated recurrent unit network to generate a joint representation tensor. The gated recurrent unit, through its internal update and reset mechanisms, adaptively determines which newly introduced damage mutation information needs to be superimposed in the current multimodal fusion feature vector, and how much of the historical damage accumulation information retained from the previous moment needs to be forgotten. Through the recursive deduction of the gated recurrent unit in the time dimension, the system aggregates isolated single-step multimodal features into a dynamic evolution sequence containing temporal context logic. The system then performs three-dimensional tensor reorganization on the hidden state vectors output by the gated recurrent unit at each time step, generating a joint representation tensor whose shape includes the time dimension, spatial feature dimension, and modal fusion channel dimension. This tensor completely and continuously encapsulates the multi-scale spatiotemporal evolution holographic information of the material under a specific loading process.

[0072] Specifically, when the segment specimen reaches the critical instability state under simulated extreme high water pressure, the three aligned local characterization vectors extracted by the system enter the fusion stage. At this time, the mechanical local characterization vector clearly indicates that it is currently in the mechanical semantics of "stress peak plateau period under high confining pressure, with macroscopic deformation sluggish". However, at this time, the CT topology characterization vector reflects that "the internal microcrack network is expanding extremely densely but slowly", and the acoustic emission characterization vector reflects that "low-frequency shear friction energy increases sharply while high-frequency tensile energy decreases". If mechanics is not used as the anchor point, the network is very likely to misjudge the dense cracks of CT as macroscopic collapse, or ignore the internal shear slip due to the lack of high-frequency signals in acoustic emission.

[0073] Through the attention mechanism of cross-modal contrastive learning, the system uses the semantics of "stress plateau period" in mechanics as a benchmark, and forces the semantics of "slow expansion" in CT and "shear friction" in acoustic emission to converge toward this benchmark to calculate the attention loss, so that the three eventually converge to the same semantic coordinate point in the latent space: "internal shear plastic damage state under high water pressure and dense constraint". After completing this precise alignment, the system splices and fuses the spatial topology information, shear energy information and plateau period mechanical information in the channel dimension to form a high-dimensional comprehensive feature vector.

[0074] The local crushing of tunnel segments under high water pressure does not occur instantaneously, but rather through a long period of slow energy accumulation in microcracks. When the comprehensive feature vector of the current moment is input into the gated loop unit, the GRU's reset gate will recognize the abnormal jump in the current "shear friction energy" and thus open the channel to receive this new feature. The update gate will combine the historical state of "slow microcrack propagation" implied in the previous moment to determine what proportion of historical accumulated damage to retain. Through this temporal context aggregation, the system's final output is not a slice of a single moment, but a joint representation tensor containing the complete evolutionary link of "early pore compaction history - mid-term slow microcrack incubation - current sudden shear friction slip". This tensor directly serves as the driving kernel of the digital twin, enabling the twin to accurately understand the extreme physical state of "apparently intact but internally damaged" brought about by high water pressure, thereby accurately simulating the sudden local crushing morphology of the tunnel segments without any warning.

[0075] refer to Figure 4 In step S13, the specific steps are as follows:

[0076] S131: Obtain a preset digital twin and load the joint characterization tensor. The joint characterization tensor performs twin simulation under the drive of the preset digital twin. At this time, the digital twin uses a parametric geometric mesh to characterize the shield tunnel segment and maps the topology and mechanical parameters in the joint characterization tensor to the initial material property field and non-uniform boundary conditions of the parametric geometric mesh.

[0077] S132: During the twin simulation, the physical constraint neural network built into the digital twin is activated. The joint representation tensor is used as input, and the forward solution is performed through the multi-layer residual network of the physical constraint neural network. The constraint conditions of the shield tunnel segment material are combined for adaptive decoding to generate a multi-physics response cloud map of the shield tunnel segment material. This multi-physics response cloud map includes a stress field distribution cloud map, a strain field evolution cloud map, and a damage variable cloud map.

[0078] In the embodiments of this application, a preset digital twin is obtained and the joint representation tensor is loaded. The joint representation tensor performs twin simulation under the drive of the preset digital twin. At this time, the digital twin uses a parametric geometric mesh to represent the shield tunnel segment, and maps the topology and mechanical parameters in the joint representation tensor to the initial material property field and non-uniform boundary conditions of the parametric geometric mesh. The topology and mechanical parameters in the joint representation tensor are mapped to the initial material property field and non-uniform boundary conditions of the parametric geometric mesh.

[0079] At this point, during the twin simulation initiation phase, the system instantiates a pre-defined digital twin of the tunnel boring machine (TBM) segment within the computational domain. To balance macroscopic geometric fidelity with microscopic damage resolution efficiency, this digital twin abandons traditional static solid modeling and employs parametric geometric meshes to spatially represent the TBM segment. Based on the standard design drawing parameters of the segment (such as width, thickness, and curvature), the system generates the external contour surface of the segment using non-uniform rational B-spline curves and fills it with an adaptive hexahedral solid mesh. The node coordinates and topological connections of this parametric geometric mesh are controlled by a set of programmable parameter vectors, enabling the mesh shape to undergo rigid translation and flexible distortion in response to externally input displacement or deformation commands. This provides a highly geometrically compatible discretized base space for subsequent multiphysics dynamic solutions.

[0080] The system receives the joint representation tensor output by the cross-modal network and uses it as a high-dimensional data engine to drive the simulation. Since the joint representation tensor is in a highly abstract latent space, the system uses a pre-trained spatial deconvolution decoder to upsample and expand the joint representation tensor along its spatial feature dimensions, accurately restoring its resolution to a three-dimensional matrix that matches the number of nodes in the parameterized geometric mesh. The system decouples and separates the channel dimensions of the decoded tensor, extracting topological feature channels that specifically represent the distribution of discontinuous surfaces and micro-defects inside the material, as well as mechanical parameter feature channels that represent the local load-bearing capacity of the material, providing a decoupled independent data source for subsequent physical property assignment.

[0081] The system performs a rigorous mapping from the data space to the physical space. For the initial material property field, the system transforms the eigenvalues ​​extracted from the topological feature channels into scalar values ​​of the local elastic modulus, Poisson's ratio, and initial damage variables at each integration point in the parameterized geometric mesh through a preset scaling transformation function. This constructs a non-uniform material stiffness matrix reflecting actual manufacturing defects on the macroscopically homogeneous pipe segment mesh. For non-uniform boundary conditions, the system combines the extracted mechanical parameter feature channels with external loading settings to transform them into distributed surface force vectors and volume force vectors of mesh surface nodes and internal interfaces. Through the above mapping operations, the internal defects and external loads, originally represented by pure mathematical tensors, are accurately translated into physical field quantities that conform to the requirements of the continuum mechanics format, thus completely completing the full physical state initialization of the digital twin before calculation.

[0082] Specifically, the system instantiated digital twin is a parametric geometric mesh model constructed at a 1:1 scale according to the actual shield tunnel segment, such as an arc segment mesh containing millions of hexahedral elements; when the system receives a joint representation tensor containing the semantics of "internal shear plastic damage under high water pressure and tight constraints", the spatial deconvolution decoder quickly unfolds it into a three-dimensional feature field matrix covering the segment mesh.

[0083] When the initial material property field is applied, the system extracts the characteristic channels representing the micro-defect topology from the tensor. Since the previous multimodal data has accurately captured the "beaded pore group" in the deep tensile side of the tube segment, the system then uses a scaling transformation function to significantly reduce the local elastic modulus of tens of thousands of grid nodes in the corresponding spatial coordinate region of the parameterized geometric grid. For example, it drops sharply from the benchmark 35 GPa to 15 GPa. At the same time, it assigns a non-zero initial damage variable to these nodes, thus "carving" out a hidden, mechanically extremely fragile, non-uniform initial material property band out of thin air inside the seemingly intact tube segment grid.

[0084] When applying non-uniform boundary conditions, the system extracts mechanical parameter channels from the joint characterization tensor and combines them with the physical settings of high water pressure geology. The system does not simply apply uniform hydrostatic pressure to the outer arc surface of the segment, but rather, based on the response characteristics of the evolution of internal defects carried in the tensor to macroscopic forces, it precisely superimposes the local gradient permeation volume force generated by the non-uniform transmission of pore water pressure and the complex spatial shear stress surface force vector caused by the misalignment of adjacent aggregate skeletons on the edge grid nodes of the aforementioned "initial material property zone".

[0085] Through the mapping in step S131 above, the initial state of the digital twin is no longer an ideal uniform pressure model, but a high-fidelity physical field model that "internally contains microcracks and weak zones confirmed by multimodal data, and externally bears non-uniform water pressure gradient loads derived from the actual evolution state." This precise initialization directly ensures that the subsequent physical constraint neural network can concentrate its computational resources on the stress concentration and shear slip evolution of the weak zone during forward solving, thereby realistically reproducing the sudden crushing mechanism of the pipe segment under high water pressure, which originates from the internal local weak zone and rapidly expands outward.

[0086] Furthermore, during the twin simulation process, the physical constraint neural network built into the digital twin is activated. Taking the joint representation tensor as input, the solution is obtained through the multi-layer residual network of the physical constraint neural network. Adaptive decoding is performed in combination with the constraint conditions of the shield tunnel segment material to generate a multi-physics response cloud map of the shield tunnel segment material. This multi-physics response cloud map includes a stress field distribution cloud map, a strain field evolution cloud map, and a damage variable cloud map.

[0087] At this point, after the digital twin completes the physical initialization of the parameterized geometric mesh, the system immediately activates the built-in physical constraint neural network to take over the subsequent solution process. This physical constraint neural network abandons the mesh stiffness matrix assembly and iterative solution of the traditional finite element method, and instead adopts a fully data-driven partial differential equation solution paradigm. The system takes the joint representation tensor containing spatiotemporal coordinate information and the initial material property field and non-uniform boundary conditions mapped in step S131, performs dimensional concatenation, and uses it as the input vector. This is then fully injected into the lowest data input layer of the physical constraint neural network at once, officially starting the mapping and solution process from high-dimensional abstract multimodal features to specific continuous physical field functions.

[0088] After the input vector is injected, the system performs forward propagation to solve the problem through a multi-layer residual network constructed inside the physical constraint neural network. This multi-layer residual network is composed of multiple cascaded residual blocks containing fully connected layers, layer normalization, and nonlinear activation functions. The cross-layer connection mechanism effectively alleviates the gradient vanishing problem when fitting high-order physical partial differential equations in deep networks. During the layer-by-layer feedforward propagation, the neurons in the hidden layer of the network continuously perform spatial transformation and manifold expansion on the nonlinear features contained in the joint representation tensor. The independent neurons in the final output layer are set as prediction functions representing the physical field state, directly outputting the displacement field prediction, stress field prediction, and damage state prediction values ​​of all spatial nodes in the parameterized geometric grid at the current time step, thereby completing the parallel feedforward inference of the global physical field at a speed of milliseconds.

[0089] After the multi-layer residual network outputs the initial physical field predictions, the system immediately introduces an adaptive decoding mechanism to ensure the physical rigor of the solution. The system embeds constitutive constraints of the shield tunnel segment concrete material under complex stress states (such as the elastoplastic damage constitutive equation based on the energy equivalence principle) and macroscopic mechanical equilibrium differential equations in the decoding layer. The system uses automatic differentiation technology to automatically calculate the first and second spatial partial derivatives of the displacement field predictions output by the network to obtain strain and strain gradient. Then, it combines the constitutive constraints to derive the theoretical stress value and calculates the difference between it and the stress field prediction value directly output by the network to generate physical residuals. Based on the convergence gradient of the physical residuals in the current training round, the system adaptively and dynamically adjusts the weight ratio of the data loss term and the physical residual penalty term in the decoding loss function, forcing the network's output distribution to strictly distort and converge to a manifold surface that conforms to objective physical laws.

[0090] After the physical residuals meet the preset convergence threshold, the system extracts the final physical field prediction value after adaptive decoding constraint correction and performs the conversion from discrete node data to continuous visualization map. The system performs mathematical reduction of the stress tensor components calculated on each node in the parameterized geometric grid to the von Mises equivalent stress or maximum principal stress, normalizes the damage prediction value according to the continuous interval from 0 to 1, and extracts the equivalent plastic strain value. According to the preset scientific computing visualization color map, such as the thermodynamic color spectrum that smoothly transitions from blue to red, the system strictly binds the above values ​​to the three-dimensional coordinates of the corresponding spatial nodes, and finally generates a multi-physics response cloud map matrix of shield tunnel segment material containing stress field distribution cloud map, strain field evolution cloud map and damage variable cloud map through pixel-by-pixel rasterization interpolation rendering.

[0091] Specifically, when the joint representation tensor of "internal hidden microcrack weak zone" and "non-uniform water pressure gradient" is input into the physical constraint neural network, the multilayer residual network begins to perform rapid feedforward inference. Due to the extremely complex stress transmission inside the pipe segment under high water pressure, the residual network, through its deep nonlinear fitting ability, quickly captures the path of how external water pressure is transmitted to the internal weak zone through the concrete skeleton in the potential space, and initially outputs a set of global displacement and stress prediction values.

[0092] Without constraints, these predicted values ​​are highly likely to violate the true physical laws under high water pressure. At this point, the adaptive decoding mechanism plays a decisive role. The system uses automatic differentiation technology to calculate the strain field by differentiating the displacement field output by the network, and verifies it by combining it with the concrete elastic-plastic damage constitutive equation under high water pressure triaxial compression. The system finds that in the preset local area of ​​"microcrack weak zone", the volumetric strain initially predicted by the network shows abnormal expansion, and the theoretical energy dissipation rate derived from this does not match the damage prediction value directly output by the network. In response to this deviation, the system immediately triggers the adaptive decoding penalty mechanism, which significantly increases the weight of the physical residual loss in this local area in backpropagation, forcing the network to modify its neuron parameters, so that the re-output damage variable must strictly obey the thermodynamic law of "shear slip causing volume expansion" under the confining pressure constraint of high water pressure.

[0093] After repeated corrections using residual penalties, the network finally outputs an exact solution that fully conforms to the physical constraints. In the final multiphysics cloud map rendering stage, the system maps this solution onto the 3D tunnel segment mesh. The generated stress field distribution cloud map clearly shows that the high water pressure load forms a uniform compressive stress layer on the surface of the tunnel segment, but at the tip of the weak zone inside, the stress lines undergo severe distortion and high concentration. The generated strain field evolution cloud map accurately presents the localized morphology of the weak zone region as it transforms from compressive strain to severe shear strain. In the generated damage variable cloud map, the originally intact areas inside the tunnel segment are presented in dark blue (damage variable < 0.1), representing safety. Only at the location of the hidden weak zone indicated by the multimodal data, a bright red damage through-line representing complete collapse (damage variable close to 0.9) erupts. The coordinated output of these three cloud maps perfectly and intuitively reproduces the entire process of the real disaster evolution of the shield tunnel segment under high water pressure, characterized by "external pressure and internal damage, local sudden collapse," in virtual space.

[0094] refer to Figure 5 In step S14, the specific steps are as follows:

[0095] S141: Obtain the measured surface displacement and strain field data of the shield tunnel segment material during the actual loading process, and determine the simulation deviation content by comparing it with the simulation data corresponding to the multiphysics response cloud map; construct dynamic penalty weight coefficients based on the simulation deviation content, calculate the adaptive weighted sum of physical constraint loss, data-driven loss and twin deviation loss, and further generate global feedback gradient;

[0096] S142: Backpropagate the global feedback gradient to the cross-modal feature decoupling network and the physical constraint neural network, and trigger the online update of the cross-modal feature decoupling network and the physical constraint neural network to dynamically adjust the projection matrix weights of the contrastive learning in the cross-modal feature decoupling network, and simultaneously correct the hidden layer parameters of the physical constraint neural network, so that the output of the digital twin converges to the real physical state of the shield tunnel segment material.

[0097] In the embodiments of this application, measured surface displacement and strain field data of shield tunnel segment material during actual loading are obtained, and the simulation deviation content is determined by comparing it with the simulation data corresponding to the multiphysics response cloud map. Based on the simulation deviation content, a dynamic penalty weight coefficient is constructed, and an adaptive weighted sum of physical constraint loss, data-driven loss and twin deviation loss is calculated to further generate a global feedback gradient. This approach takes into account the overall consideration of the simulation deviation content and ensures the accuracy of the dynamic penalty weight coefficient.

[0098] At this point, after the digital twin outputs the multiphysics response cloud map, the system immediately introduces the measured feedback data of the shield tunnel segment material during the actual loading physical process to perform closed-loop verification. The system acquires the full-field surface displacement data collected by the three-dimensional digital image correlation technology arranged on the surface of the segment, as well as the high-precision surface strain time series data collected by the distributed fiber optic grating sensor. In view of the difference between the measured data and the simulation cloud map in terms of spatial resolution and sampling frequency, the system uses the surface node coordinates of the parameterized geometric grid of the digital twin as the reference target point, and uses bicubic spline interpolation to spatially resample the measured surface displacement field. Then, through low-pass filtering and time axis alignment, the measured strain field is mapped to the corresponding discrete time step of the simulation solution, thereby generating the measured physical field matrix that absolutely coincides with the multiphysics response cloud map in spatiotemporal coordinates.

[0099] After unifying the spatiotemporal reference, the system performs point-by-point and component-by-component difference calculations on the simulated displacement vectors and strain tensors of the surface nodes extracted from the simulation cloud map and the measured physical field matrix after resampling. The system not only calculates the global root mean square error to assess the overall offset, but also focuses on constructing a deviation identification operator based on local spatial gradients. By calculating the local variance and extreme value distribution of the strain field deviation, the system accurately isolates the overall rigid translation error caused by the distortion of the mesh boundary conditions and the local strain concentration error caused by the inaccuracy of the constitutive model parameters. The system structurally encapsulates the spatial coordinates, deviation amplitude, and divergence rate over time of the above-mentioned local strain extreme value deviation regions, and clearly defines them as the current simulation deviation content of the digital twin, providing accurate spatial positioning and quantitative indicators for subsequent targeted correction.

[0100] After locking in the simulation deviation content, the system enters the adaptive reconstruction stage of the loss function to prevent gradient conflicts in multi-objective optimization. Based on the local deviation magnitude and divergence rate in the determined simulation deviation content, the system constructs a dynamic penalty weight coefficient matrix. At this time, for local grid regions where the deviation magnitude exceeds the preset safety threshold and exhibits a rapid divergence state, the system dynamically increases the corresponding physical constraint penalty weight coefficient to strongly constrain the prediction behavior that violates physical laws in that region. At the same time, for macroscopic deviations such as overall rigid translation error, the system proportionally increases the penalty weight coefficient of data-driven loss. This dynamic penalty weight coefficient can nonlinearly and adaptively decay or surge according to the spatiotemporal evolution characteristics of the simulation deviation content in each load step, ensuring that the penalty focus of the loss function is always accurately locked on the dimension where the twin deviates most severely from the real physical state.

[0101] Under the modulation of dynamic penalty weight coefficients, the system calculates an adaptive weighted sum of physical constraint loss, data-driven loss, and twin bias loss. The physical constraint loss consists of the residuals of partial differential equations and material constitutive equations, the data-driven loss consists of the residuals of measured surface displacement and strain, and the twin bias loss consists of a special penalty term specifically constructed for the local bias content determined in the second stage of step S141. The system multiplies the above three types of losses by their corresponding dynamic penalty weight coefficients and sums them to generate a comprehensive loss scalar. Relying on the automatic differentiation engine of the deep learning framework, the system performs chain rule differentiation on this comprehensive loss scalar, backtracks along the computational graph of the built-in physical constraint neural network and cross-modal feature decoupling network of the digital twin, and finally aggregates to generate a global feedback gradient tensor that can simultaneously correct physical solution bias and low-level feature representation bias.

[0102] Specifically, when the digital twin outputs the stress and strain field cloud maps of the local crushing of the tunnel segment, the system simultaneously acquires the actual displacement field captured by the DIC system and the surface strain data obtained by FBG on the surface of the tunnel segment during the high water pressure loading test. The system extrapolates the simulation cloud map to the surface of the tunnel segment and performs node-by-node comparison with the measured data. Through the deviation identification operator, the system keenly discovers that: on the macroscopic pressure surface of the tunnel segment, the simulated displacement and the measured displacement are highly consistent; however, in the surface projection area corresponding to the preset "micro-crack weak zone" inside the tunnel segment, the strain field output by the simulation presents an extremely sharp narrow band-shaped strain set, while the measured surface strain field also shows concentration, but the strain gradient is relatively gentle and the peak value only stays at 3200 micro-strain. Based on this, the system determines that the simulation deviation is: "There is a non-physical false strain localization over-evolution phenomenon in the weak zone area."

[0103] In response to simulation deviations, the system immediately triggers a dynamic penalty weight coefficient construction mechanism. Since the deviation amplitude in this region is huge and would cause the neural network solution to collapse if not suppressed, the system instantly amplifies the penalty weight coefficients for "non-uniform energy dissipation rate" and "abnormal damage evolution gradient" in the physical constraint loss by tens of times within the spatial coordinate grid corresponding to the "micro-crack weak zone," forcibly suppressing the network's tendency to output extreme values ​​in this region. At the same time, the weight of the overall data-driven loss is appropriately reduced to prevent the model from disrupting the internal physical balance in order to accommodate the small displacement fluctuations on the surface.

[0104] The system adaptively weights and sums the amplified physical constraint loss, the stable data-driven loss, and the twin bias loss specifically designed for spurious localization. The system uses an automatic differential engine to penetrate the physical constraint neural network in reverse and calculates a global feedback gradient that can forcibly correct the weights of neurons in the weak band network and flatten spurious strain gradients. This gradient flow not only corrects the physical solution bias of PINN backward but also propagates forward to the cross-modal network in step S121, fine-tuning the attention allocation of the acoustic emission radio frequency band energy characteristics. Thus, in the next simulation time step, it completely eliminates the "stress lock-up" and "spurious crushing" artifacts caused by over-coupling of modal features in the high water pressure simulation, forcing the twin to strictly converge to the real material softening path under high water pressure.

[0105] Furthermore, the global feedback gradient is backpropagated to the cross-modal feature decoupling network and the physical constraint neural network, triggering online updates of the cross-modal feature decoupling network and the physical constraint neural network. This dynamically adjusts the projection matrix weights of the contrastive learning in the cross-modal feature decoupling network and simultaneously corrects the hidden layer parameters of the physical constraint neural network, so that the output of the digital twin converges to the true physical state of the shield tunnel segment material. This achieves a twin simulation of the joint representation tensor, fully considering the simulation deviation content, the cross-modal feature decoupling network and the physical constraint neural network, and improving the accuracy of the true physical state of the shield tunnel segment material.

[0106] At this point, after obtaining the global feedback gradient generated by the comprehensive multidimensional loss, the system immediately initiates a deep backpropagation mechanism across heterogeneous computing modules. Starting from the output layer of the physical constraint neural network built into the digital twin, the system uses the chain rule to backtrack the gradient field layer by layer according to the directed acyclic topology of the computation graph. When the gradient flow backtracks to the input end of the physical constraint neural network, the system uses the Jacobian transpose operation of matrix calculus to seamlessly penetrate the modality mapping boundary of the global feedback gradient of the physical domain and accurately project it into the neuron connection links inside the upstream cross-modal feature decoupling network, thereby constructing a continuous, differentiable end-to-end gradient backpropagation channel from the final physical field deviation to the underlying data feature representation.

[0107] After the global feedback gradient is injected into the cross-modal feature decoupling network, the system performs online parameter updates for the modal semantic alignment module. The system identifies specific gradient components in the gradient flow that are mapped to the modal orthogonal projection module. These components carry the correction direction for the "current feature fusion does not fully reflect the real physical damage mechanism". The system uses an adaptive moment estimation optimizer to dynamically adjust the weight parameters of the projection matrix in contrastive learning based on this gradient component. At this point, the system reduces the weights of projection matrices that cause "excessive overlap or erroneous rejection" of heterogeneous modal features in the latent space, and increases the weights that can strengthen the clustering of features related to the real physical state. This changes the projection trajectory of subsequent multimodal data on the orthogonal basis, realizing the self-evolution of the underlying data semantic alignment strategy.

[0108] While dynamically adjusting the feature decoupling network, the system performs synchronous physical field solution correction on the physically constrained neural network. The system distributes the partial differential equation residual gradient components left inside the physically constrained neural network in the global feedback gradient to each hidden layer in the multi-layer residual network. The system corrects the weight matrix and bias vector of the fully connected neurons in the hidden layer through gradient descent, forcibly changing the nonlinear spatial transformation logic of the network for the joint representation tensor. This correction process enables the network hidden layer to "forget" the non-physical fitting patterns that caused the previous simulation deviations, relearn and solidify the differential operator approximation that conforms to the constitutive relations of real materials and the law of conservation of momentum, thereby achieving a leap in solution accuracy without changing the network topology.

[0109] After completing the synchronous online update of the projection matrix weights and hidden layer parameters, the system enters the convergence closed-loop verification stage. The system uses the updated dual-network architecture to re-execute the forward propagation calculation, generate a new round of multiphysics response cloud map, and again call the measured surface displacement field and strain field data for residual comparison. The system monitors the change curves of the global physical residual and the measured root mean square error in real time. If the error index shows a monotonically decreasing trend and is lower than the preset physical tolerance threshold, it is determined that the output of the digital twin has eliminated model drift and successfully converged to the true physical state of the shield tunnel segment material. If the error still oscillates, the system will convert the new round of error into gradient again and continue to drive the dual network iteration until the twin and the physical entity achieve high-fidelity state synchronization.

[0110] Optionally, when the global feedback gradient is backpropagated to the cross-modal feature decoupling network and the physically constrained neural network, the system triggers an online update mechanism for both networks. Specifically, the system uses an adaptive moment estimation optimizer to update the parameters, where the learning rate is initially set to 0.001, the exponential decay rate of the first-order moment estimation is set to 0.9, the exponential decay rate of the second-order moment estimation is set to 0.999, and the numerical stability constant is set to 10 to the power of negative 8. The above parameters remain consistent in each backpropagation and do not automatically decay with training rounds to ensure the stability of the online update process.

[0111] Regarding update conditions, the system triggers network updates only under the following three conditions: First, the overall loss scalar of the current simulation increases by more than 10% compared to the loss value at the previous moment; second, the root mean square error between the measured surface displacement field and the simulated displacement field is greater than a preset threshold for two consecutive time steps, which is set to 0.05 mm for the displacement field and 50 microstrains for the strain field; third, among the simulation deviations determined by step S141, the local deviation amplitude exceeds a preset safety threshold and the deviation divergence rate shows a positive increasing trend. In terms of update frequency, the system adopts a periodic update strategy instead of updating every step. That is, after completing the simulation derivation for five consecutive time steps, the system summarizes the global feedback gradients of these five time steps, calculates their arithmetic mean, and then performs a unified network parameter update. This design can avoid parameter oscillations caused by single-step random noise.

[0112] Specifically, the system has identified the fatal deviation of "excessive concentration of simulated strain in the weak zone region of microcracks deviating from the actual measurement" through step S141, and generated a global feedback gradient carrying a forced strain gradient flattening command. At this time, the system starts cross-network backpropagation. This gradient flow washes over the hidden layer of the physical constraint neural network. The system synchronously corrects the parameters of the hidden layer according to the gradient, forcing the residual block inside PINN to "erase" the irrational nonlinear activation response that was previously over-amplified to conform to the topology of the weak zone, so that it returns to the real elastoplastic damage mechanical trajectory under high water pressure triaxial compression.

[0113] More importantly, this gradient flow penetrates PINN and reaches the upstream cross-modal feature decoupling network. The system found that the previous contrastive learning projection matrix overemphasized the weight of "sharp edges of pores" in the CT topology features when fusing data, while ignoring the suppressive factor of "high water pressure dense friction" reflected in the acoustic emission radio frequency band energy features. Therefore, based on the feedback gradient, the system dynamically lowers the weight coefficient of the corresponding sharp topology features of CT in the projection matrix, while increasing the weight coefficient of the corresponding low-frequency friction energy features of acoustic emission.

[0114] After the above-mentioned online updates via dual networks, the semantic cognition of the system was reconstructed: it no longer regards the weak zone as an "absolute fracture cavity," but rather as a "dense fracture zone compacted by high water pressure but undergoing shear slip." Driven by this revised logic, the system performs forward solving again. In the newly generated multiphysics response cloud map, the strain concentration zone in this region is significantly wider and the gradient is gentler, with the peak value accurately falling back to 3250 microstrain. Moreover, the stress field distribution perfectly matches the displacement field measured by DIC. The system's judgment error is below the threshold, strictly converging to the true physical catastrophe evolution state under complex geological conditions of high water pressure.

[0115] Please see Figure 6 The twin simulation system for the multimodal data of the tunnel boring machine segment material is applied to the aforementioned twin simulation method for the multimodal data of the tunnel boring machine segment material; the twin simulation system for the multimodal data of the tunnel boring machine segment material includes:

[0116] The multimodal data module 21 is used to acquire X-ray CT voxel data of shield tunnel segment material during the preparation stage, and combine it with the acoustic emission waveform data of shield tunnel segment and the mechanical response data of shield tunnel segment material for spatiotemporal alignment, thereby determining the multimodal data of shield tunnel segment material;

[0117] The joint representation tensor module 22 is used to input multimodal data into the cross-modal feature decoupling network, extract the local representation vectors of each mode respectively, and perform semantic alignment on the local representation vectors, thereby fusing them to generate the joint representation tensor of the shield tunnel segment material;

[0118] The twin simulation module 23 is used to input the joint representation tensor into the preset digital twin and trigger the twin simulation of the joint representation tensor. Using the joint representation tensor as the boundary condition, the physical constraint neural network built into the digital twin performs forward solution to generate the corresponding multiphysics response cloud map.

[0119] The online update module 24 is used to extract key damage features from the multiphysics response cloud map, combine them with the measured data of the shield tunnel segment material to determine the simulation deviation content, and trigger the online update of the cross-modal feature decoupling network and the physical constraint neural network, so that the output of the digital twin converges to the real physical state of the shield tunnel segment material.

[0120] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0121] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and embodiments are to be considered exemplary only.

[0122] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A twin simulation method for multimodal data of tunnel segment materials, characterized in that, include: X-ray CT voxel data of shield tunnel segment materials during the preparation stage were obtained, and spatiotemporal alignment was performed by combining the acoustic emission waveform data of shield tunnel segments and the mechanical response data of shield tunnel segment materials to determine the multimodal data of shield tunnel segment materials. Multimodal data is input into a cross-modal feature decoupling network, and local representation vectors of each modality are extracted. The local representation vectors are then semantically aligned to generate a joint representation tensor of the shield tunnel segment material. The joint representation tensor is input into a preset digital twin, and the twin simulation of the joint representation tensor is triggered. The joint representation tensor is used as the boundary condition, and the physical constraint neural network built into the digital twin performs a forward solution to generate the corresponding multiphysics response cloud map. Key damage features are extracted from the multiphysics response cloud map, and the simulation deviation is determined by combining the measured data of the shield tunnel segment material. The online update of the cross-modal feature decoupling network and the physical constraint neural network is triggered, so that the output of the digital twin converges to the real physical state of the shield tunnel segment material.

2. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 1, characterized in that, The process involves acquiring X-ray CT voxel data of the shield tunnel segment material during the preparation stage, and combining this data with the acoustic emission waveform data and mechanical response data of the shield tunnel segment material for spatiotemporal alignment to determine the multimodal data of the shield tunnel segment material, including: X-ray CT voxel data, acoustic emission waveform data, and mechanical response data of shield tunnel segment materials during the preparation stage are acquired; the X-ray CT voxel data are denoised and reconstructed to extract the corresponding topological features; wavelet packet decomposition is performed on the acoustic emission waveform data to extract frequency band energy features; and the mechanical response data are identified to extract the corresponding stress-strain data.

3. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 2, characterized in that, The process of acquiring X-ray CT voxel data of the shield tunnel segment material during the preparation stage, and combining it with the acoustic emission waveform data and mechanical response data of the shield tunnel segment material for spatiotemporal alignment to determine the multimodal data of the shield tunnel segment material, further includes: Mechanical response data is synchronized with X-ray CT voxel data and acoustic emission waveform data through time axis interpolation. Microscopic topological features, frequency band energy features and stress-strain data are jointly registered to eliminate the spatiotemporal scale differences and phase delays between multimodal data, generating multimodal data of shield tunnel segment materials. This multimodal data presents spatial coordinate alignment, time series synchronization and physical state correlation.

4. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 1, characterized in that, The process of inputting multimodal data into a cross-modal feature decoupling network, extracting local representation vectors for each modality, and semantically aligning these local representation vectors to generate a joint representation tensor for the tunnel segment material includes: Multimodal data is input into a cross-modal feature decoupling network, which includes a 3D convolutional twin branch, a temporal graph convolutional branch, and a fully connected feature mapping branch that process X-ray CT voxel data, acoustic emission waveform data, and mechanical response data, respectively. The corresponding local representation vectors are extracted through each branch, and a modal orthogonal projection module is introduced to perform semantic alignment on each local representation vector to output the corresponding common semantic features.

5. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 4, characterized in that, The step of inputting multimodal data into a cross-modal feature decoupling network, extracting local representation vectors for each modality, and semantically aligning the local representation vectors to fuse and generate a joint representation tensor for the tunnel segment material further includes: Using the local representation vectors corresponding to the mechanical response data as semantic anchors, the semantic alignment loss is determined through the attention mechanism of cross-modal contrastive learning, so as to minimize the distance between the local representation vectors of different modalities representing the same physical damage state in the latent feature space. The common semantic features are spliced ​​and fused, and then the temporal context is aggregated through a gated recurrent unit to generate a joint representation tensor of the shield tunnel segment material.

6. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 1, characterized in that, The process involves inputting the joint representation tensor into a preset digital twin and triggering a twin simulation of the joint representation tensor. Using the joint representation tensor as boundary conditions, the physical constraint neural network built into the digital twin performs a forward solution to generate the corresponding multiphysics response cloud map, including: A preset digital twin is obtained and the joint representation tensor is loaded. The joint representation tensor performs twin simulation under the drive of the preset digital twin. At this time, the digital twin uses a parametric geometric mesh to represent the shield tunnel segment and maps the topology and mechanical parameters in the joint representation tensor to the initial material property field and non-uniform boundary conditions of the parametric geometric mesh.

7. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 6, characterized in that, The step of inputting the joint representation tensor into a preset digital twin and triggering a twin simulation of the joint representation tensor, using the joint representation tensor as boundary conditions, and having the physical constraint neural network built into the digital twin perform a forward solution to generate the corresponding multiphysics response cloud map, further includes: During the twin simulation, the physical constraint neural network built into the digital twin is activated. The joint representation tensor is used as input, and the forward solution is performed through the multi-layer residual network of the physical constraint neural network. The constraint conditions of the shield tunnel segment material are combined for adaptive decoding to generate a multi-physics response cloud map of the shield tunnel segment material. This multi-physics response cloud map includes a stress field distribution cloud map, a strain field evolution cloud map, and a damage variable cloud map.

8. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 1, characterized in that, The key damage features extracted from the multiphysics response cloud map are combined with measured data of the tunnel lining segment material to determine the simulation deviation, and online updates of the cross-modal feature decoupling network and physical constraint neural network are triggered, so that the output of the digital twin converges to the true physical state of the tunnel lining segment material, including: The measured surface displacement and strain field data of the shield tunnel segment material under actual loading are obtained, and the simulation deviation content is determined by comparing it with the simulation data corresponding to the multiphysics response cloud map. Based on the simulation deviation content, a dynamic penalty weight coefficient is constructed, and the adaptive weighted sum of physical constraint loss, data-driven loss and twin deviation loss is calculated to further generate a global feedback gradient.

9. The twin simulation method for multimodal data of shield tunnel segment materials according to claim 8, characterized in that, The process of extracting key damage features from the multiphysics response cloud map, combining them with measured data of the tunnel lining segment material to determine the simulation deviation, and triggering online updates of the cross-modal feature decoupling network and the physical constraint neural network, so that the output of the digital twin converges to the true physical state of the tunnel lining segment material, also includes: The global feedback gradient is backpropagated to the cross-modal feature decoupling network and the physical constraint neural network, triggering online updates of the cross-modal feature decoupling network and the physical constraint neural network. This dynamically adjusts the projection matrix weights of the contrastive learning in the cross-modal feature decoupling network and simultaneously corrects the hidden layer parameters of the physical constraint neural network, so that the output of the digital twin converges to the true physical state of the shield tunnel segment material.

10. A twin simulation system for multimodal data of tunnel segment materials, characterized in that, The twin simulation system for the multimodal data of the shield tunnel segment material is applied to the twin simulation method for the multimodal data of the shield tunnel segment material as described in any one of claims 1-9; the twin simulation system for the multimodal data of the shield tunnel segment material includes: The multimodal data module is used to acquire X-ray CT voxel data of shield tunnel segment materials during the preparation stage, and combine it with the acoustic emission waveform data of shield tunnel segments and the mechanical response data of shield tunnel segment materials for spatiotemporal alignment, thereby determining the multimodal data of shield tunnel segment materials; The joint representation tensor module is used to input multimodal data into the cross-modal feature decoupling network, extract the local representation vectors of each mode, and perform semantic alignment on the local representation vectors, thereby fusing them to generate a joint representation tensor of the shield tunnel segment material. The twin simulation module is used to input the joint representation tensor into a preset digital twin and trigger the twin simulation of the joint representation tensor. Using the joint representation tensor as boundary conditions, the physical constraint neural network built into the digital twin performs forward solving to generate the corresponding multiphysics response cloud map. The online update module is used to extract key damage features from the multiphysics response cloud map, combine them with the measured data of the shield tunnel segment material to determine the simulation deviation, and trigger the online update of the cross-modal feature decoupling network and the physical constraint neural network, so that the output of the digital twin converges to the real physical state of the shield tunnel segment material.

Images