Shield tunneling ahead unfavorable stratum identification method based on elastic wave coherence imaging and physical information graph neural network

Abstract

The application discloses a shield front adverse stratum advanced identification method based on elastic wave coherent imaging and a physical information graph neural network, and comprises the following steps: collecting full waveform time series data sets of a shield tunneling machine; decoupling wave field components of mixed vibration signals according to the full waveform time series data sets and a space-time coupling tensor model, and acquiring pure stratum characteristic source signals; extracting multi-dimensional waveform features and constructing a waveform feature set according to the pure stratum characteristic source signals; performing wave field coherent imaging and feature matching according to the waveform feature set and the purified pre-contact reflection wave signals, and determining the spatial distribution of an adverse stratum interface; and predicting the type, spatial distribution and risk level of the adverse stratum in front of a shield tunneling face according to the spatial distribution through a mechanical endogenous driving graph wave field neural network model. The application can solve the core problems of the prior art, such as single scene, disconnection between identification and landing and no engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of intelligent construction and advanced prediction of adverse geological conditions in underground engineering shield tunnels, and particularly relates to an advanced identification method for adverse strata ahead of shield tunnels based on elastic wave coherent imaging and physical information graph neural networks. Background Technology

[0002] With its advantages of safety, efficiency, and minimal impact on the surrounding environment, shield tunneling has become a core construction method for urban underground space development and inter-regional water conveyance tunnel construction. During the tunneling process, the type and physical and mechanical properties of the strata at the tunnel face directly determine the matching of tunneling parameters, construction safety management, and equipment wear control. Therefore, accurate and real-time identification of the strata in front of the tunnel machine is a core prerequisite for realizing intelligent shield tunneling.

[0003] Currently, the technology for identifying and predicting geological formations during tunnel boring machine (TBM) construction suffers from several insurmountable technical defects: Pre-construction geological drilling techniques are hampered by scattered drilling points and an inability to cover the entire tunnel's strata, failing to meet the real-time identification requirements of dynamic construction; Advanced geological prediction technologies during construction, including advanced drilling, ground-penetrating radar, and seismic wave detection, suffer from drawbacks such as time constraints, high detection costs, significant susceptibility to electromagnetic and mechanical interference, and a high dependence on personnel experience for accuracy, making continuous and timely strata identification throughout the entire construction cycle difficult; Strata identification technology based on TBM tunneling parameters, while existing research utilizes construction parameters such as total thrust of the TBM, cutterhead torque, and propulsion speed for strata identification, these parameters represent a delayed response after the cutterhead cuts through the strata, failing to anticipate changes in the strata ahead. Furthermore, they are highly susceptible to human intervention in adjusting construction parameters, failing to accurately reflect the inherent mechanical properties of the strata, resulting in low accuracy in identifying unfavorable strata and high false alarm / missed alarm rates. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for advanced identification of unfavorable geological formations ahead of tunnel boring machines (TBMs) based on elastic wave coherent imaging and physical information graph neural networks. This method can solve the core problems of existing technologies, such as limited application scenarios, disconnect between identification and implementation, and lack of engineering application value. It has significant engineering implications for realizing intelligent TBM tunneling, preventing construction risks, and improving tunneling efficiency.

[0005] To achieve the above objectives, this invention provides a method for advance identification of unfavorable geological formations ahead of a tunnel boring machine based on elastic wave coherent imaging and a physical information graph neural network, comprising: Collect the full waveform time-series dataset of the tunnel boring machine; Based on the full waveform time series dataset and the spatiotemporal coupling tensor model, the wave field components of the hybrid vibration signal are decoupled to obtain the pure stratum characteristic source signal. Based on the pure formation characteristic source signal, multi-dimensional waveform features are extracted and a waveform feature set is constructed. Based on the waveform feature set and the purified pre-contact reflected wave signal, wavefield coherent imaging and feature matching are performed to determine the spatial distribution of unfavorable stratigraphic interfaces. Based on the spatial distribution, the type, spatial distribution, and risk level of unfavorable strata ahead of the tunnel boring machine are predicted using a mechanically endogenous driven graph wave field neural network model.

[0006] Optionally, the full waveform time-series dataset of the tunnel boring machine can be collected, including: Based on the vibration acceleration sensor arrays arranged along different cutting radii of the tunnel boring machine cutterhead, the triaxial vibration acceleration waveform signals during the cutting process of the cutterhead are collected in real time. Based on the cutter head azimuth encoder, real-time rotation azimuth data of the cutter head is collected synchronously. The original tunneling parameters of the tunnel boring machine are collected synchronously according to the PLC control system. According to the PTP high-precision time synchronization protocol, the triaxial vibration acceleration waveform signal, the real-time rotation azimuth data of the cutterhead, and the native tunneling parameters are time-aligned to construct the full waveform time series dataset.

[0007] Optionally, based on the full waveform time-series dataset and the spatiotemporal coupled tensor model, wavefield component decoupling is performed on the hybrid vibration signal to obtain pure formation characteristic source signals, including: Based on hobbing cutting mechanics, cutter head structure dynamics and elastic wave propagation theory, a time-varying and space-varying vibration signal spatiotemporal coupling tensor model is constructed. Based on the spatiotemporal coupling tensor model and the intrinsic constraints of cutting mechanics, initialize the hybrid transfer matrix; Based on the initialized hybrid transfer matrix, the framed hybrid signal tensor is decomposed using a joint diagonalization algorithm to obtain the optimal estimate of the hybrid matrix, thus completing the separation of multiple vibration sources. Based on the optimal estimate, the multi-source vibration source separation result is reconstructed; Based on the mechanical characteristics and wave field propagation characteristics of various pre-calibrated source signals, the multi-source vibration source separation results are classified, and the pre-contact reflection wave source signals that only reflect the unfavorable strata ahead are screened and purified as the pure strata characteristic source signals.

[0008] Optionally, the spatiotemporal coupling tensor model is: X(t,θ,r)=A(t,θ,r)·S(t,θ,r)+N(t,θ,r); Where X(t,θ,r) is the tensor of the hybrid vibration signal collected at time t, cutterhead azimuth angle θ, and cutting radius r; A(t,θ,r) is the time-varying and space-varying hybrid transfer matrix; S(t,θ,r) is the source signal matrix, which includes four independent source signals: pure formation characteristic source signal, multi-roller coupled vibration source, equipment inherent vibration source, and control disturbance vibration source; N(t,θ,r) is the random noise matrix.

[0009] Optionally, extracting multi-dimensional waveform features and constructing a waveform feature set based on the pure formation feature source signal includes: Based on the pure formation characteristic source signal, extract time domain features, frequency domain features, time-frequency domain features, pre-contact weak signal features, impedance abrupt change features specific to unfavorable formations, and reflected wave coherence features to obtain characteristic parameters; The feature parameters are standardized to construct the waveform feature set.

[0010] Optionally, based on the waveform feature set and the purified pre-contact reflected wave signal, wavefield coherent imaging and feature matching are performed to determine the spatial distribution of undesirable stratigraphic interfaces, including: Based on the spatial location and synchronous sampling sequence of the multi-sensor unit, the purified pre-contact reflected wave signal is subjected to time delay compensation and coherent superposition to construct the stratum impedance profile in front of the tunnel face. Based on the waveform feature set and the pre-constructed waveform feature library specific to unfavorable formations, feature similarity is calculated using a similarity matching algorithm; Based on the stratigraphic impedance profile and characteristic similarity calculation results, the spatial distribution morphology, dip angle, scale, and differences in mechanical parameters of the two sides of the adverse stratigraphic interface are inverted using the travel-time tomography algorithm to determine the spatial distribution.

[0011] Optionally, the mechanically intrinsically driven graph wave field neural network model includes: an input layer, a graph wave field convolution module, a dedicated attention module, and an output module; The input layer is used to construct a feature map structure with stratigraphic spatial topological relationships based on the feature sequence of continuous multi-rings, the wavefield coherent imaging profile of the corresponding rings, the excavated geological verification data, and the sequence of tunneling working condition parameters. The graph wave field convolution module is used to update node features and learn features based on the feature graph structure through a graph wave field convolution module that intrinsically embeds the formation elastic wave wave equation and the mechanical constitutive equation. The dedicated attention module is used to effectively enhance the signal based on the learned node characteristics through the dedicated attention module for weak signals in poor formations. The output module is used to output the adverse geological conditions ahead of the shield tunneling face, the spatial distribution of the interface, its scale and depth, the range of geological mechanical parameters, and the adverse geological risk level, based on the enhanced features.

[0012] Optionally, node feature updating and feature learning include: ; In the formula, h v l+1 Let d be the feature vector of node v in layer (l+1), and N(v) be the set of neighboring nodes of node v. v d u Let G(r) be the degree of the node. vu ,ω) is the graph convolution kernel function derived from the Green's function of the elastic wave equation, r vu Let ω be the spatial distance between nodes u and v, and W be the elastic wave angular frequency. l W0 l Let Φ(h) be a learnable weight matrix. v l ,h u l ) is an intrinsic constraint term in mechanics, f( ) is a non-linear activation function. Let u be the feature vector of node v's neighbor node u in the l-th layer of the network. This represents the original feature vector of node v before it is updated in the l-th layer of the network.

[0013] Optionally, the training of the mechanically endogenous driving graph wave field neural network model includes: Based on indoor roller cutting test data, formation elastic wave propagation numerical simulation data, and typical adverse formation mechanical parameter datasets, the network was pre-trained with mechanical rules in an unsupervised manner, and the core parameters of the mechanical endogenous graph wave field convolution module were frozen. Based on the measured data of the shield tunneling starting section, geological survey data, and adverse strata calibration data, the typical adverse strata samples are weighted and enhanced, and the attention module and output module of the network are fine-tuned with small learning rate increments.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: 1. Comprehensive data governance: This invention uses adaptive spectral subtraction and wavelet threshold denoising technology to accurately remove non-stratum interference components. The signal-to-noise ratio of the decoupled pure stratum feature waveform is improved by more than 82%, which solves the problem of interference signals overwhelming stratum features, significantly improves data quality and modeling stability, and effectively addresses the problem of inconsistent data noise and quality in engineering field. 2. Self-learning and scalability: The model in this invention can be dynamically updated with tunneling data. Combined with the dynamic time warping algorithm, it achieves accurate feature matching and has the ability to continuously evolve and migrate across projects. It solves the technical problem that existing methods use static model parameters and are difficult to adapt to changes in working conditions at different stages. 3. System-level integration: This invention can be directly embedded into the control system to achieve closed-loop linkage between stratum identification and parameter control, improving construction safety and automation levels, and solving the technical problem of low closed-loop integration between existing methods and shield tunneling control systems. The modification cost is extremely low, possessing strong engineering applicability and large-scale promotion value. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the method for advance identification of unfavorable strata ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the layout of the three-ring differential distributed vibration acceleration sensor array according to an embodiment of the present invention; Figure 3 This is a diagram of the neural network PIGWNN construction framework according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the spatiotemporal coherence imaging principle of pre-contact wavefield in poor formations according to an embodiment of the present invention; Among them, 1. hobbing cutter, 2. vibration acceleration sensor. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] This embodiment proposes a method for advance identification of adverse geological formations ahead of a tunnel boring machine (TBM) based on elastic wave coherent imaging and a physical information graph neural network. Figure 1 As shown, the specific steps include: Collect the full waveform time-series dataset of the tunnel boring machine; Based on the full waveform time series dataset and the spatiotemporal coupling tensor model, the wave field components of the hybrid vibration signal are decoupled to obtain the pure stratum characteristic source signal. Among them, the hybrid vibration signal is a composite vibration signal of the shield cutterhead from multiple sources after being quantized by spatiotemporal coupling tensor, according to the spatiotemporal coupling tensor model described below: X(t,θ,r)=A(t,θ,r)·S(t,θ,r)+N(t,θ,r); The main components include: 1. Pure formation characteristic source signal: the elastic wave excited by the cutting of the formation by the cutter 1 propagates in front of the tunnel face, and is reflected when it encounters an unfavorable formation interface, forming a pre-contact reflection wave.

[0019] 2. Multi-roller coupled vibration source: Dozens of roll cutters on the cutterhead cut the formation simultaneously, and the cutting vibrations generated by each cutterhead are coupled and superimposed through the cutterhead spoke structure to form a composite vibration.

[0020] 3. Inherent vibration sources of the equipment: The inherent vibrations generated by the operation of the tunnel boring machine's own mechanical system mainly include the rotation of the main bearing, the operation of the hydraulic pump station, and the extension and retraction of the propulsion cylinder.

[0021] 4. Controlling the source of disturbance vibration: Transient vibrations caused by the dynamic adjustment of tunneling parameters by the shield tunneling control system, including adjustment of propulsion speed, regulation of soil chamber pressure, and changes in cutterhead speed.

[0022] These are transient disturbances caused by human intervention and need to be identified and eliminated through temporal correlation and mechanical characteristics.

[0023] Based on the pure formation characteristic source signal, multi-dimensional waveform features are extracted and a waveform feature set is constructed. Based on the waveform feature set and the purified pre-contact reflected wave signal, wavefield coherent imaging and feature matching are performed to determine the spatial distribution of unfavorable stratigraphic interfaces. Based on the spatial distribution, the type, spatial distribution, and risk level of unfavorable strata ahead of the tunnel boring machine are predicted using a mechanically endogenous driven graph wave field neural network model.

[0024] Specifically, this embodiment is based on the collaborative execution of a shield machine cutterhead vibration acceleration sensor array, cutterhead azimuth encoder, PLC control system, and onboard edge computing terminal. The core components include full-space synchronous waveform data acquisition, wave field component decoupling embedded with multi-field coupled cutting mechanics, multi-dimensional feature extraction for advanced identification of unfavorable strata, feature matching and spatial tomography localization of unfavorable strata interfaces, advanced prediction of unfavorable strata ahead based on wave field imaging and PIGWNN, and engineering closed-loop control of the identification results. The sensor deployment scheme for data acquisition is as follows: Figure 2 As shown, the model framework for the neural network prediction stage is as follows: Figure 3 As shown, the principle of the wavefield coherent imaging stage is as follows: Figure 4 As shown.

[0025] Furthermore, the full waveform time-series dataset of the tunnel boring machine includes: Based on the vibration acceleration sensor arrays arranged along different cutting radii of the tunnel boring machine cutterhead, the triaxial vibration acceleration waveform signals during the cutting process of the cutterhead are collected in real time. Based on the cutter head azimuth encoder, real-time rotation azimuth data of the cutter head is collected synchronously. The original tunneling parameters of the tunnel boring machine are collected synchronously according to the PLC control system. According to the PTP high-precision time synchronization protocol, the triaxial vibration acceleration waveform signal, the real-time rotation azimuth data of the cutterhead, and the native tunneling parameters are time-aligned to construct the full waveform time series dataset.

[0026] Specifically, S1, full-space synchronous waveform data acquisition: The triaxial vibration acceleration waveform signal during the cutting process of the cutterhead is acquired in real time through a vibration acceleration sensor array deployed along different cutting radii of the tunnel boring machine cutterhead; the real-time rotation azimuth data of the cutterhead is acquired synchronously through the cutterhead azimuth encoder; the original tunneling parameters of the tunnel boring machine are acquired synchronously through the PLC control system; the data is time-aligned through the PTP high-precision time synchronization protocol, and the onboard edge computing terminal constructs a full waveform time series dataset; The sensor array is configured to satisfy the elastic wave spatial sampling theorem, ensuring that reflected waves from unfavorable strata ahead are sampled without aliasing. Sensor units are arranged along the three sets of different cutting radii of the tunnel boring machine cutterhead, corresponding to the central cutterhead group, the front cutterhead group, and the edge cutterhead group, respectively. Four triaxial vibration acceleration sensors 2 are arranged at equal intervals along the circumference of the cutterhead spokes on each cutting radius. The sensors are installed on the cutterhead spokes next to the corresponding cutterhead holder by means of bolt fixing and welding reinforcement. The cutter head spindle azimuth encoder and sensor array are equipped with a phase synchronization triggering mechanism to achieve coherent sampling of reflected waves for the hobbing cutting vibration source. When the cutter head rotates to the abrupt change area of ​​the interface of the identified unfavorable strata, the hardware automatically increases the sensor sampling frequency. When the cutter head rotates to the low-feature area of ​​the homogeneous strata, the hardware automatically decreases the sampling frequency, realizing the coordinated optimization of high-precision sampling in the unfavorable strata feature area and computing power saving in the low-feature area.

[0027] Furthermore, based on the full waveform time-series dataset and the spatiotemporal coupled tensor model, wavefield component decoupling is performed on the hybrid vibration signal to obtain pure formation characteristic source signals, including: Based on hobbing cutting mechanics, cutter head structure dynamics and elastic wave propagation theory, a time-varying and space-varying vibration signal spatiotemporal coupling tensor model is constructed. Based on the spatiotemporal coupling tensor model and the intrinsic constraints of cutting mechanics, initialize the hybrid transfer matrix; Based on the initialized hybrid transfer matrix, the framed hybrid signal tensor is decomposed using a joint diagonalization algorithm to obtain the optimal estimate of the hybrid matrix. Based on the optimal estimate, the multi-source vibration source separation result is reconstructed; Based on the mechanical characteristics and wave field propagation characteristics of various pre-calibrated source signals, the multi-source vibration source separation results are classified, and the pre-contact reflection wave source signals that only reflect the unfavorable strata ahead are screened and purified as the pure strata characteristic source signals.

[0028] Specifically, S2, waveform decoupling and preprocessing of multi-field coupled cutting mechanics: The airborne edge computing terminal constructs a time-varying and space-varying vibration signal spatiotemporal coupled tensor model based on the hob-cutter head-surrounding rock coupled cutting mechanics model and elastic wave propagation theory. Through the blind source separation algorithm of cutting mechanics intrinsic constraints, the collected mixed vibration signal is fully decoupled, and the multi-hob coupling vibration components, the equipment's inherent vibration components, and the control disturbance components are separated in one go.

[0029] Optimal estimated route: I. Joint diagonalization decomposition → obtaining the optimal estimate of the mixture matrix: Based on the initialized hybrid transfer matrix, a fourth-order cumulant matrix is ​​constructed from the framed hybrid signal tensor. Signal correlation is eliminated by the joint diagonalization algorithm to obtain Â(t,θ,r).

[0030] II. Based on the optimal estimate Â, the multi-source vibration source separation result S^ is obtained through reconstruction: By substituting the optimal estimate  into the spatiotemporal coupling tensor model through tensor operations, the inverse transformation is completed: the operation outputs the multi-source vibration source separation result, which decomposes the mixed vibration signal into four independent source signals.

[0031] III. Based on the separation result S^, pure formation characteristic source signals are obtained through classification and purification: Taking the reconstructed S^(t,θ,r) as the processing object, and combining the pre-calibrated mechanical characteristics and wave field propagation characteristics: 1. Use Hilbert-Huang transform to extract the marginal spectrum and travel time features of each source signal; 2. Eliminate three types of interference signals: multi-roller coupling, inherent equipment, and control disturbance vibration sources; 3. Screening and purifying the pre-contact reflection source signals that only reflect the unfavorable strata ahead, thus obtaining the final pure strata characteristic source signals as described in the patent specification.

[0032] Furthermore, in step S2, the spatiotemporal coupling tensor model is: X(t,θ,r)=A(t,θ,r)·S(t,θ,r)+N(t,θ,r); Where X(t,θ,r) is the tensor of the hybrid vibration signal collected at time t, cutterhead azimuth angle θ, and cutting radius r; A(t,θ,r) is the time-varying and space-varying hybrid transfer matrix, which characterizes the spatiotemporal transmission characteristics of multi-source vibration on the cutterhead; S(t,θ,r) is the source signal matrix, which includes four independent source signals: pure stratum characteristic source signal, multi-roller coupled vibration source, equipment inherent vibration source, and control disturbance vibration source; N(t,θ,r) is the random noise matrix.

[0033] The specific execution steps are as follows: S2.1 Initialization of the hybrid matrix under the intrinsic constraints of cutting mechanics: Based on hob cutting mechanics, cutter head structure dynamics and elastic wave propagation theory, the a priori constraint conditions of the hybrid matrix A are derived. The hob vibration transfer function with the same cutting radius has circumferential periodicity, and the non-Gaussianity of the formation cutting source signal and the reflected wave source signal is significantly stronger than that of the interference source signal. The hybrid matrix is ​​initialized based on these a priori conditions. S2.2 Spatiotemporal Coupled Tensor Joint Diagonalization Decomposition: To address the non-stationary characteristics of the rotating cutting head, the mixed signal tensor is divided into frames, with each 15° rotation of the cutting head as a frame, to ensure the quasi-stationarity of the signal within a single frame; a fourth-order cumulant matrix is ​​constructed for each frame signal, and the optimal estimate A of the mixed matrix is ​​solved by the joint diagonalization algorithm to achieve multi-source vibration source separation; S2.3 Source signal classification and purification of reflection waves from unfavorable formations: Based on the pre-calibrated mechanical characteristics and wave field propagation characteristic labels of various source signals, the marginal spectral characteristics and travel time characteristics of each separated source signal are extracted by Hilbert-Huang transform. Support vector machine is used to classify the source signals, accurately screen and purify the pre-contact reflection wave source signals that only reflect the unfavorable formations ahead. S2.4 Adaptive Resolution Optimization for Non-Stationary Signals: Based on the real-time rotational speed and azimuth angle of the cutter head, the time window length of the signal frame is adaptively adjusted to process the cutting vibration and reflected wave signals, adapting to the non-stationary signal characteristics brought about by rotary cutting.

[0034] Furthermore, based on the pure formation feature source signal, extracting multi-dimensional waveform features and constructing a waveform feature set includes: Based on the pure formation characteristic source signal, extract time domain features, frequency domain features, time-frequency domain features, pre-contact weak signal features, impedance abrupt change features specific to unfavorable formations, and reflected wave coherence features to obtain characteristic parameters; The feature parameters are standardized to construct the waveform feature set.

[0035] Specifically, S3, multi-dimensional feature extraction for adaptive advanced identification: The airborne edge computing terminal performs multi-dimensional feature extraction on pure stratum feature source signals, including time domain features, frequency domain features, time-frequency domain features, pre-contact weak signal features, impedance change features specific to unfavorable strata, and reflected wave coherence features. All feature parameters are standardized using the Z-score method to construct a waveform feature set for adaptive advanced identification.

[0036] Time-frequency domain feature extraction: The pure formation feature waveform is decomposed into three layers of wavelet packets using the db10 wavelet basis to obtain the wavelet packet energy coefficients of eight frequency bands. The time-frequency domain energy feature vector is constructed to reflect the waveform abrupt change characteristics of the adverse formation interface. Pre-contact weak signal characteristics: Extract the elastic wave reflection signal characteristics of micro-fractures in unexcavated strata ahead of the tunnel face and the interfaces of unfavorable strata, providing a basis for advanced identification; Abrupt impedance change characteristic specific to unfavorable formations: Designed to address the strong impedance difference between unfavorable formations and intact surrounding rock, the expression is F. imp =ΔA cos(θ) / Δt; In the formula, ΔA is the abrupt change in the amplitude of the reflected wave, Δt is the travel time difference of the reflected wave, and θ is the dip angle of the stratum interface.

[0037] Furthermore, based on the waveform feature set and the purified pre-contact reflected wave signal, wavefield coherent imaging and feature matching are performed to determine the spatial distribution of unfavorable stratigraphic interfaces, including: Based on the spatial location and synchronous sampling sequence of the multi-sensor unit, the purified pre-contact reflected wave signal is subjected to time delay compensation and coherent superposition to construct the stratum impedance profile in front of the tunnel face. Based on the waveform feature set and the pre-constructed waveform feature library specific to unfavorable formations, feature similarity is calculated using a similarity matching algorithm; Based on the stratigraphic impedance profile and characteristic similarity calculation results, the spatial distribution morphology, dip angle, scale, and differences in mechanical parameters of the two sides of the adverse stratigraphic interface are inverted using the travel-time tomography algorithm to determine the spatial distribution.

[0038] Specifically, S4, stratigraphic interface feature matching and spatial positioning: The airborne edge computing terminal first performs spatiotemporal coherent imaging of the pre-contact wavefield of the unfavorable strata based on the purified pre-contact reflection wave signal. Then, combined with feature matching and travel-time tomography inversion, it completes the accurate spatial positioning and parameter inversion of the interface of the unfavorable strata. The specific steps are as follows: S4.1, Spatiotemporal coherent imaging of pre-contact wavefield: Based on the spatial position and synchronous sampling sequence of multiple sensing units, the reflected wave signal is time-delayed and coherently superimposed to construct the wave impedance profile of the strata in front of the tunnel face. By using the pre-calibrated wave impedance abrupt change threshold of the undesirable strata, the spatial distribution range, interface position and burial depth of the undesirable strata are initially identified. S4.2, Matching the Feature Library of Adverse Strata: The real-time standardized feature set is matched with a pre-constructed waveform feature library specifically for adverse strata. The feature library is constructed through indoor roller cutting tests and pre-calibration of field engineering geological data. For common adverse geological types in shield tunneling, corresponding wave impedance features, waveform features, and mechanical parameter labels are collected to form a standard feature library. Feature similarity matching uses a dynamic time warping algorithm to calculate the similarity between the real-time feature sequence and the standard sequence. The preset similarity threshold is 0.85. When the similarity exceeds the threshold, it is determined to be an adverse stratum of the corresponding type. S4.3 Spatial Tomographic Location of Unfavorable Strata: Combining coherent imaging results and feature matching results, the travel-time tomography algorithm is used to invert and calculate the spatial distribution, dip angle, scale, and mechanical parameter differences of the interface of unfavorable strata within the tunnel face, thereby completing the accurate spatial location and parameter inversion of unfavorable strata.

[0039] Furthermore, the mechanically intrinsically driven graph wave field neural network model includes: an input layer, a graph wave field convolution module, a dedicated attention module, and an output module; The input layer is used to construct a feature map structure with stratigraphic spatial topological relationships based on the feature sequence of continuous multi-rings, the wavefield coherent imaging profile of the corresponding rings, the excavated geological verification data, and the sequence of tunneling working condition parameters. The graph wave field convolution module is used to update node features and learn features based on the feature graph structure through a graph wave field convolution module that intrinsically embeds the formation elastic wave wave equation and the mechanical constitutive equation. The dedicated attention module is used to effectively enhance the signal based on the learned node characteristics through the dedicated attention module for weak signals in poor formations. The output module is used to output the adverse geological conditions ahead of the shield tunneling face, the spatial distribution of the interface, its scale and depth, the range of geological mechanical parameters, and the adverse geological risk level, based on the enhanced features.

[0040] Specifically, S5, Leading identification of unfavorable formations based on wavefield imaging and neural networks: Based on the wavefield coherent imaging results of S4, the accurate advance prediction of adverse strata ahead of the tunnel boring machine is achieved through the mechanically endogenous driven graph wavefield neural network (PIGWNN) model.

[0041] The airborne edge computing terminal takes a feature sequence of five consecutive rings, wavefield coherent imaging profiles of the corresponding rings, excavated geological verification data of the corresponding rings, and tunneling condition parameter sequences as model inputs. It first constructs a feature map structure with strata spatial topology, and then completes feature learning through a graph wavefield convolution module that embeds strata mechanical constitutive and elastic wave equations. It enhances effective signals through a dedicated attention module for weak signals in adverse strata, and finally outputs the adverse strata conditions in front of the shield tunneling face, the spatial distribution of the interface, its scale and depth, the range of strata mechanical parameters, and the adverse geological risk level.

[0042] Furthermore, the mechanically endogenously driven Graph Wave Field Neural Network (PIGWNN) model comprises, in sequence: Input Layer: The spatiotemporal wavefield feature map structure construction module constructs a stratigraphic feature map structure with spatial topological relationships from the feature set of the continuous tunneling ring and the wavefield imaging profile. In this structure, the graph nodes correspond to the spatial positions of the sensing units within the tunnel face, and the node features are the feature vectors of the corresponding positions and the wavefield imaging data. The spatial coordinates of the nodes are uniquely determined by the sensor array installation position and the cutterhead azimuth angle. The graph edge weights are endogenously determined by the path difference of elastic wave propagation and the attenuation law of the formation medium. Core layer: The mechanically endogenous graph wave field neural network embeds the formation elastic wave equation and mechanical constitutive equation into the design of the graph convolution kernel function. The graph convolution kernel function is directly derived from the Green's function of the elastic wave equation, and the node feature update rule strictly follows the elastic wave propagation law and the formation mechanical constitutive model. ; In the formula, h v l+1 Let d be the feature vector of node v in layer (l+1), and N(v) be the set of neighboring nodes of node v. v d u Let G(r) be the degree of the node. vu ,ω) is the graph convolution kernel function derived from the Green's function of the elastic wave equation, r vu Let ω be the spatial distance between nodes u and v, and W be the elastic wave angular frequency. l W0 l Let Φ(h) be a learnable weight matrix. v l ,h u l ) is an intrinsic constraint term in mechanics, f( ) is a non-linear activation function. Let u be the feature vector of node v's neighbor node u in the l-th layer of the network. This represents the original feature vector of node v before it is updated in the l-th layer of the network.

[0043] Furthermore, the training of the mechanically endogenous driving graph wave field neural network model includes: Based on indoor roller cutting test data, formation elastic wave propagation numerical simulation data, and typical adverse formation mechanical parameter datasets, the network was pre-trained with mechanical rules in an unsupervised manner, and the core parameters of the mechanical endogenous graph wave field convolution module were frozen. Based on the measured data of the shield tunneling starting section, geological survey data, and adverse strata calibration data, the typical adverse strata samples are weighted and enhanced, and the attention module and output module of the network are fine-tuned with small learning rate increments.

[0044] Specifically, the training of the PIGWNN model is divided into two dedicated stages: Unsupervised pre-training of mechanical rules: Based on indoor roller cutting test data, formation elastic wave propagation numerical simulation data, and typical adverse formation mechanical parameter datasets, the network is pre-trained in an unsupervised manner, freezing the core parameters of the mechanical endogenous graph wave field convolution module to ensure that the network feature learning strictly follows the laws of formation mechanics and elastic wave propagation theory throughout the entire process. The incremental fine-tuning of adverse strata sample enhancement is based on measured data, geological survey data, and adverse strata calibration data from the shield tunneling starting section. Typical adverse strata samples are weighted and enhanced, and the attention layer and output layer of the network are fine-tuned with small learning rate increments. This ensures the model's generalization ability under different engineering scenarios and different adverse strata types, and significantly reduces the false alarm and false negative rates of adverse strata.

[0045] A lightweight, layered inference architecture is set up for edge and cloud collaboration: the sensor's built-in MCU is responsible for lightweight computations such as signal framing and preprocessing, reducing data transmission volume; the airborne edge terminal deploys wave field decoupling algorithm for forward inference, feature extraction and advance prediction inference of PIGWNN model, and compresses the network parameter size to less than 5MB through model quantization technology, with a single-loop data full-process inference time of ≤30ms, meeting the real-time requirements of the project.

[0046] S6. Engineered closed-loop control of adverse formation identification results: Tiered early warning: For the identified unfavorable strata and risk level, the corresponding level of audible and visual early warning is triggered, a pop-up notification is displayed on the screen in the shield tunneling operator's room, and the information is simultaneously pushed to the project department's remote control platform. If a red level one early warning is triggered, the emergency response plan is simultaneously pushed to the platform. Model adaptive update: Every 20 rings of tunneling, the system supplements and improves the feature library for undesirable strata based on the geological sketches, slag removal analysis, and borehole verification data of the already excavated sections, fine-tunes the PIGWNN model with a small learning rate increment, freezes the weights of the bottom mechanical convolutional layers, and only updates the parameters of the fully connected layers to ensure the stable identification accuracy of undesirable strata during long-distance tunneling.

[0047] This embodiment is applied to a shield tunneling project with a length of 1860m. The shield tunneling section traverses complex geological formations, including silty clay, moderately weathered limestone, slightly weathered limestone, composite strata with soft upper and hard lower layers, and fault fracture zones. The unfavorable geological sections account for 32% of the total. During construction, accidents such as cutter head breakage, cutterhead jamming, and water and mud inrush at the tunnel face are very likely to occur. Therefore, the requirements for the advanced identification accuracy and real-time performance of the geological interface ahead are extremely high.

[0048] System hardware deployment, installation, and debugging implementation steps: The hardware system in this embodiment is deployed on the left-line tunnel boring machine (TBM) and is installed and debugged during the TBM assembly phase. The specific implementation steps are as follows: Deployment and installation of a three-ring differential distributed vibration acceleration sensor array: Layout Design: For the cutterhead structure of this tunnel boring machine, sensor arrays are arranged along three concentric circles with different cutting radii, as detailed below: Central ring: Corresponding to the central hob set, with a cutting radius of 0.7m, four triaxial vibration acceleration sensors are evenly spaced along the circumference of the central spokes of the cutter head, with a circumferential spacing of 90° between adjacent sensors; Front ring: Corresponding to the front hob set, with a cutting radius of 2.3m, four triaxial vibration acceleration sensors are evenly spaced along the circumference of the front spokes of the cutter head, with a circumferential spacing of 90° between adjacent sensors; Edge ring: corresponding to the edge hobbing cutter group, with a cutting radius of 3.1m, four triaxial vibration acceleration sensors are evenly spaced along the circumferential direction of the spokes on the edge of the cutter head, with a circumferential spacing of 90° between adjacent sensors; the circumferential spatial spacing between adjacent sensors in the same group is 0.36m~1.57m, which is less than 1 / 2 of the minimum wavelength of elastic waves (3.2m) in the limestone strata of this project, satisfying the elastic wave spatial sampling theorem.

[0049] Furthermore, M8 high-strength bolts are used to fix the sensor mounting base to the cutterhead spokes, with a bolt tightening torque ≥25N・m. The base is then fully reinforced with welding around it, with a weld leg height ≥6mm, to prevent the sensor from falling off during the tunneling process.

[0050] Azimuth encoder installation and synchronous triggering debugging: Encoder installation: Install the absolute photoelectric encoder coaxially with the cutter head spindle via a coupling, ensuring that the encoder rotation is completely synchronized with the cutter head rotation, and align the encoder zero point with the vertically upward position of the cutter head to complete the angle zero point calibration; Hardware synchronous trigger debugging: Connect the synchronous trigger pulse output interface of the encoder to the trigger input interface of 12 sensors via shielded cable. Set the encoder to output one synchronous trigger pulse for every 1° rotation, so as to realize hardware-level synchronous sampling of all sensors and encoder, with a sampling time synchronization error ≤1μs.

[0051] Implementation steps for system calibration and model pre-training before tunnel boring machine launch: The training of the PIGWNN model is divided into two dedicated phases: Unsupervised pre-training of mechanical rules: Based on indoor roller cutting test data, formation elastic wave propagation numerical simulation data, and typical adverse formation mechanical parameter datasets, the network is pre-trained in an unsupervised manner, freezing the core parameters of the mechanical endogenous graph wave field convolution module to ensure that the network feature learning strictly follows the laws of formation mechanics and elastic wave propagation theory throughout the entire process. The incremental fine-tuning of adverse stratum sample enhancement is based on measured data, geological survey data, and adverse stratum calibration data from the shield tunneling starting section. A difficult case sample enhancement strategy is adopted to weight and enhance typical adverse stratum samples. The attention layer and output layer of the network are fine-tuned with a small learning rate increment, and the weights of the mechanical endogenous convolutional layer are frozen to reduce the false alarm and false negative rates of adverse stratum.

[0052] Furthermore, a lightweight, layered inference architecture is set up for edge-cloud collaboration: the sensor's built-in MCU is responsible for lightweight computations such as signal framing and preprocessing, reducing data transmission volume; the airborne edge terminal deploys wave field decoupling algorithm for forward inference, and PIGWNN model for feature extraction and advance prediction inference. Through model quantization technology, the network parameter size is compressed to less than 5MB, and the single-loop data full-process inference time is ≤30ms, meeting the real-time requirements of the project; the project's cloud platform is responsible for model pre-training, big data incremental updates and accuracy verification, and distributes the updated lightweight model to the edge terminal.

[0053] The algorithm implementation steps throughout the shield tunneling process are as follows: S1 full-space synchronous waveform data acquisition implementation: When the tunnel boring machine reached the 723rd ring, the cutterhead rotation speed was set to 1.2 r / min, the advance speed was set to 20 mm / min, and the penetration depth was 8 mm / r. The system performed the following data acquisition operations: Twelve triaxial vibration acceleration sensors, with a basic sampling frequency of 1024Hz, acquire real-time full waveform signals of X, Y, and Z axis vibration acceleration during the process of the hob cutting the formation; The cutter head azimuth encoder synchronously collects the real-time rotation azimuth data of the cutter head and triggers pulses through hardware synchronization to ensure that all sensor sampling is completely synchronized with the cutter head azimuth. The PLC control system synchronously collects native tunneling parameters such as total thrust, cutterhead torque, propulsion speed, penetration depth, cutterhead rotation speed, soil chamber pressure, and propulsion cylinder chamber pressure at a frequency of 10Hz. All data are aligned to nanosecond-level timing using the PTP high-precision time synchronization protocol. The airborne edge computing terminal matches the corresponding cutterhead cutting space position, tunneling ring number and working condition parameters for each set of waveform data, and constructs a full waveform time series dataset with spatial coordinate labels. The system detected a significant abrupt change in the characteristics of the reflected wave and automatically triggered a high-frequency sampling mode, increasing the sensor sampling frequency to 4096Hz to perform high-precision sampling of the abrupt change region.

[0054] S2 wavefield component decoupling and preprocessing implementation: The airborne edge computing terminal performs full decoupling and preprocessing on the full waveform data collected from ring 723. The specific steps are as follows: Based on the pre-constructed coupled cutting mechanics model of hob, cutter head, and surrounding rock, a time-varying and space-varying vibration signal spatiotemporal coupled tensor model is constructed. Based on the inherent vibration characteristic spectrum of the equipment and the cutting mechanics parameters of the hob calibrated before the start, the initialization of the hybrid matrix is ​​completed. The mixed signal tensor is divided into frames, with each frame representing a 15° rotation of the cutter head. The single-loop data is divided into 24 frames. A fourth-order cumulant matrix is ​​constructed for each frame signal. The optimal estimate of the mixed matrix is ​​obtained by using a joint diagonalization algorithm, thereby achieving spatial domain separation of multiple vibration sources. By extracting the marginal spectral features and travel time features of each separated source signal through Hilbert-Huang transform, and using a pre-trained support vector machine to classify the source signals, the interference components such as cutting direct wave, equipment natural vibration, and multi-roller coupled vibration are accurately separated, and the pre-contact reflection wave source signal that only reflects the unfavorable strata ahead is obtained. Based on the real-time rotation speed of the cutter head of 1.2 r / min, the time window length of the signal framing is adaptively adjusted to 208 ms to achieve variable resolution processing of non-stationary signals.

[0055] S3 Multidimensional Feature Extraction Implementation: The airborne edge computing terminal performs feature extraction on the purified cutting direct wave source signal and the pre-contact reflected wave source signal: Time-frequency domain feature extraction: The db10 wavelet basis is used to perform 3-level wavelet packet decomposition on each frame of signal to obtain wavelet packet energy coefficients in 8 frequency bands, and a time-frequency domain energy feature vector is constructed. Extraction of pre-contact weak signal characteristics, abrupt changes in impedance specific to unfavorable formations, and coherent characteristics of reflected waves: Extract the corresponding feature parameters according to the aforementioned formulas and rules of this invention; All feature parameters are standardized using the Z-score method to construct the waveform feature set of the 723rd ring, which is then output to subsequent modules.

[0056] Implementation of S4 poor stratigraphic coherence imaging and spatial tomography localization: Spatiotemporal coherent imaging of the pre-contact wavefield: Based on the spatial installation coordinates and synchronous sampling sequence of 12 sensors, the purified reflected wave signals are subjected to time delay compensation and phase correction. The delay superposition algorithm is used to complete the coherent superposition of multiple signals and construct a two-dimensional profile of the strata wave impedance in the range of 3-5 rings in front of the tunnel face. By using the pre-calibrated threshold for adverse strata wave impedance abrupt change, multiple wave impedance abrupt changes in the strata ahead are preliminarily identified, corresponding to the interfaces of adverse strata and karst development areas. Matching of unfavorable formation feature library: The standardized feature set of ring 723 is matched with the initialized unfavorable formation feature library. The similarity between the real-time feature sequence and the standard sequence is calculated using the dynamic time warping algorithm. The similarity is greater than 0.85. Spatial tomography localization: Combining coherent imaging results and feature matching results, the travel-time tomography algorithm is used to invert and calculate the spatial distribution morphology, strike, dip angle, and development scale of the interface of undesirable strata, as well as the mechanical parameters of the strata on both sides of the interface, thus completing the accurate spatial localization of undesirable strata.

[0057] S5 Advance Prediction Implementation of Unfavorable Formations Ahead of S5: The airborne edge computing terminal takes the feature sequence of five consecutive rings (719-723), the wavefield coherent imaging profile of the corresponding rings, the excavated geological verification data of the corresponding rings, and the tunneling condition parameter sequence as input. Through the PIGWNN model deployed on the edge terminal, it completes forward inference and outputs accurate prediction results of unfavorable strata in rings 3-5 ahead of the shield tunneling face, as follows: The first ring ahead: a composite stratum with soft upper layer and hard lower layer, consisting of silty clay in the upper part and moderately weathered limestone in the lower part. The strata interface dips at 30° and is located on the left side of the working face. The identification confidence level is 99.2%, and the adverse geological risk level is III. The first two rings: a composite stratum with soft upper layer and hard lower layer, with moderately weathered limestone accounting for 75%, identification confidence level of 98.8%, and adverse geological risk level III; The area in front of the third ring consists of moderately weathered limestone with localized karst development. The cave measures 0.8m × 1.2m, is partially filled, and is located in the lower right part of the working face. The uniaxial compressive strength of the strata is 60-75MPa, with an identification confidence level of 98.7%. The adverse geological risk level is II. The area in the first four rings is moderately weathered limestone, a complete rock mass with a uniaxial compressive strength of 65-80 MPa. There are no unfavorable strata developments, with an identification confidence level of 98.2% and an unfavorable geological risk level of I. The area ahead, ring 5, is an interface between moderately and slightly weathered limestone, with a locally developed fault fracture zone. The fracture zone is approximately 2.2m wide, with an identification confidence level of 97.8%, and is classified as a Level III adverse geological risk.

[0058] During the model inference process, all prediction results are verified through a mechanical back-verification layer, and finally the prediction results are pushed to the closed-loop control module.

[0059] S6 Engineering Closed-Loop Control Implementation: Tiered early warning: For the identified unfavorable strata and risk level, the corresponding level of audible and visual early warning is triggered, a pop-up notification is displayed on the screen in the shield tunneling operator's room, and the information is simultaneously pushed to the project department's remote control platform. If a red level one early warning is triggered, the emergency response plan is simultaneously pushed to the platform. Model adaptive update: Every 20 rings of tunneling, the system supplements and improves the feature library for undesirable strata based on the geological sketches, slag removal analysis, and borehole verification data of the already excavated sections, fine-tunes the PIGWNN model with a small learning rate increment, freezes the weights of the bottom mechanical convolutional layers, and only updates the parameters of the fully connected layers to ensure the stable identification accuracy of undesirable strata during long-distance tunneling.

[0060] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for advance identification of unfavorable geological formations ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network, characterized in that, include: Collect the full waveform time-series dataset of the tunnel boring machine; Based on the full waveform time series dataset and the spatiotemporal coupling tensor model, the wave field components of the hybrid vibration signal are decoupled to obtain the pure stratum characteristic source signal. Based on the pure formation characteristic source signal, multi-dimensional waveform features are extracted and a waveform feature set is constructed. Based on the waveform feature set and the purified pre-contact reflected wave signal, wavefield coherent imaging and feature matching are performed to determine the spatial distribution of unfavorable stratigraphic interfaces. Based on the spatial distribution, the type, spatial distribution, and risk level of unfavorable strata ahead of the tunnel boring machine are predicted using a mechanically endogenous driven graph wave field neural network model.

2. The method for advance identification of unfavorable strata ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 1, characterized in that, The full waveform time series dataset of the tunnel boring machine includes: Based on the vibration acceleration sensor arrays arranged along different cutting radii of the tunnel boring machine cutterhead, the triaxial vibration acceleration waveform signals during the cutting process of the cutterhead are collected in real time. Based on the cutter head azimuth encoder, real-time rotation azimuth data of the cutter head is collected synchronously. The original tunneling parameters of the tunnel boring machine are collected synchronously according to the PLC control system. According to the PTP high-precision time synchronization protocol, the triaxial vibration acceleration waveform signal, the real-time rotation azimuth data of the cutterhead, and the native tunneling parameters are time-aligned to construct the full waveform time series dataset.

3. The method for advance identification of unfavorable geological formations ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 1, characterized in that, Based on the full waveform time-series dataset and the spatiotemporal coupling tensor model, wave field component decoupling is performed on the hybrid vibration signal to obtain pure formation characteristic source signals, including: Based on hobbing cutting mechanics, cutter head structure dynamics and elastic wave propagation theory, a time-varying and space-varying vibration signal spatiotemporal coupling tensor model is constructed. Based on the spatiotemporal coupling tensor model and the intrinsic constraints of cutting mechanics, initialize the hybrid transfer matrix; Based on the initialized hybrid transfer matrix, the framed hybrid signal tensor is decomposed using a joint diagonalization algorithm to obtain the optimal estimate of the hybrid matrix. Based on the optimal estimate, the multi-source vibration source separation result is reconstructed; Based on the mechanical characteristics and wave field propagation characteristics of various pre-calibrated source signals, the multi-source vibration source separation results are classified, and the pre-contact reflection wave source signals that only reflect the unfavorable strata ahead are screened and purified as the pure strata characteristic source signals.

4. The method for advance identification of unfavorable strata ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 3, characterized in that, The spatiotemporal coupling tensor model is as follows: X(t,θ,r)=A(t,θ,r)·S(t,θ,r)+N(t,θ,r); Where X(t,θ,r) is the tensor of the hybrid vibration signal collected at time t, cutterhead azimuth angle θ, and cutting radius r; A(t,θ,r) is the time-varying and space-varying hybrid transfer matrix; S(t,θ,r) is the source signal matrix, which includes four independent source signals: pure formation characteristic source signal, multi-roller coupled vibration source, equipment inherent vibration source, and control disturbance vibration source; N(t,θ,r) is the random noise matrix.

5. The method for advance identification of unfavorable strata ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 1, characterized in that, Based on the pure formation feature source signal, extracting multi-dimensional waveform features and constructing a waveform feature set includes: Based on the pure formation characteristic source signal, extract time domain features, frequency domain features, time-frequency domain features, pre-contact weak signal features, impedance abrupt change features specific to unfavorable formations, and reflected wave coherence features to obtain characteristic parameters; The feature parameters are standardized to construct the waveform feature set.

6. The method for advance identification of unfavorable strata ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 1, characterized in that, Based on the waveform feature set and the purified pre-contact reflection wave signal, wavefield coherent imaging and feature matching are performed to determine the spatial distribution of unfavorable stratigraphic interfaces, including: Based on the spatial location and synchronous sampling sequence of the multi-sensor unit, the purified pre-contact reflected wave signal is subjected to time delay compensation and coherent superposition to construct the stratum impedance profile in front of the tunnel face. Based on the waveform feature set and the pre-constructed waveform feature library specific to unfavorable formations, feature similarity is calculated using a similarity matching algorithm; Based on the stratigraphic impedance profile and characteristic similarity calculation results, the spatial distribution morphology, dip angle, scale, and differences in mechanical parameters of the two sides of the adverse stratigraphic interface are inverted using the travel-time tomography algorithm to determine the spatial distribution.

7. The method for advance identification of unfavorable geological formations ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 1, characterized in that, The mechanically intrinsically driven graph wave field neural network model includes: an input layer, a graph wave field convolution module, a dedicated attention module, and an output module; The input layer is used to construct a feature map structure with stratigraphic spatial topological relationships based on the feature sequence of continuous multi-rings, the wavefield coherent imaging profile of the corresponding rings, the excavated geological verification data, and the sequence of tunneling working condition parameters. The graph wave field convolution module is used to update node features and learn features based on the feature graph structure through a graph wave field convolution module that intrinsically embeds the formation elastic wave wave equation and the mechanical constitutive equation. The dedicated attention module is used to effectively enhance the signal based on the learned node characteristics through the dedicated attention module for weak signals in poor formations. The output module is used to output the adverse geological conditions ahead of the shield tunneling face, the spatial distribution of the interface, its scale and depth, the range of geological mechanical parameters, and the adverse geological risk level, based on the enhanced features.

8. The method for advance identification of unfavorable strata ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 7, characterized in that, Node feature updating and feature learning include: ； In the formula, h v l+1 Let d be the feature vector of node v in layer (l+1), and N(v) be the set of neighboring nodes of node v. v d u Let G(r) be the degree of the node. vu ,ω) is the graph convolution kernel function derived from the Green's function of the elastic wave equation, r vu Let ω be the spatial distance between nodes u and v, and W be the elastic wave angular frequency. l W0 l Let Φ(h) be a learnable weight matrix. v l ,h u l ) is an intrinsic constraint term in mechanics, f( ) is a non-linear activation function. Let u be the feature vector of node v's neighbor node u in the l-th layer of the network. This represents the original feature vector of node v before it is updated in the l-th layer of the network.

9. The method for advance identification of unfavorable strata ahead of a tunnel boring machine based on elastic wave coherent imaging and physical information graph neural network according to claim 7, characterized in that, The training of the mechanically endogenous driving graph wave field neural network model includes: Based on indoor roller cutting test data, formation elastic wave propagation numerical simulation data, and typical adverse formation mechanical parameter datasets, the network was pre-trained with mechanical rules in an unsupervised manner, and the core parameters of the mechanical endogenous graph wave field convolution module were frozen. Based on the measured data of the shield tunneling starting section, geological survey data, and adverse strata calibration data, the typical adverse strata samples are weighted and enhanced, and the attention module and output module of the network are fine-tuned with small learning rate increments.

Images