Tunneling target positioning method based on complex loose body multi-modal data survey

Abstract

The application discloses a tunneling target positioning method based on complex loose body multi-modal data survey, relates to the technical field of underground space artificial intelligence survey, and comprises the following steps: synchronously collecting multi-modal sensing data and tunneling working condition data in a tunneling process and performing adaptive noise reduction processing; extracting an energy envelope feature sequence of the multi-modal sensing data and performing depth mapping alignment; inputting the aligned multi-modal data features into a double-branch feature coding network, introducing a semantic prior vector as a context constraint, distributing fusion weights of the multi-modal data features through a cross attention mechanism, and outputting a high-dimensional fusion feature tensor; and constructing a physical information constraint neural network, taking electromagnetic and acoustic physical equations and auxiliary sensor data as constraint terms of a loss function for calculation, and outputting three-dimensional positioning coordinates of a tunneling target in a tunnel global coordinate system. The method eliminates time-depth conversion errors caused by complex loose body multi-variation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence surveying technology in underground space, and in particular to a method for locating tunneling targets based on multimodal data surveying of complex loose bodies. Background Technology

[0002] When tunnel boring machines (TBMs) and shield tunneling equipment traverse complex loose strata such as water-rich faults, glacial till, and strongly weathered rock, they are highly susceptible to engineering disasters such as boulder collisions, water and mud inrushes, and face instability. Therefore, achieving accurate spatial positioning of adverse geological formations and high-risk targets ahead of the tunneling operation, and conducting advanced geological forecasting, has become a core key to ensuring the safety and efficiency of the project. To overcome the limitations of penetration depth or resolution of single physical detection methods, related surveying technologies have gradually evolved from traditional single-source detection to collaborative detection using multi-modal sensor data such as ground-penetrating radar and seismic waves.

[0003] However, for complex loose bodies, a unique geological object, existing multimodal exploration and positioning techniques rely on empirically set fixed wave velocities or simplified layered models for time-depth conversion, leading to misalignment of different modal data mappings in the spatial domain. Regarding multimodal data fusion mechanisms, existing methods generally employ rigid feature stitching or fixed-weight decision-level fusion, ignoring the dynamic response differences of different physical fields under specific geological semantic environments. Furthermore, in target localization inversion models, existing inversion networks are mostly purely data-driven. When dealing with loose bodies with extremely complex and variable physical properties, they often encounter severe homospectral heterogeneity. Due to the lack of rigid boundary constraints based on the underlying physical propagation and evolution laws, as well as real-time calibration of in-situ sensor data, the output 3D positioning coordinates are prone to deviating from the target. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a tunneling target location method based on multimodal data survey of complex loose bodies to solve the prominent problems of existing multimodal geological survey and target location technologies, such as large spatiotemporal alignment errors of heterogeneous signals, lack of semantic adaptive adjustment mechanism for cross-modal fusion, and easy susceptibility to physical ambiguity based on pure data-driven inversion.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for locating tunneling targets based on multimodal data survey of complex loose bodies, which includes simultaneously acquiring multimodal sensing data and tunneling condition data during the tunneling process and performing adaptive noise reduction processing; Extract the energy envelope feature sequence from the multimodal sensing data, solve the optimal matching path using the dynamic time warping algorithm, and perform depth mapping alignment; The aligned multimodal data features are input into a dual-branch feature encoding network. A semantic prior vector is introduced as a context constraint. The fusion weights of the multimodal data features are adaptively allocated through a cross-attention mechanism, and a high-dimensional fusion feature tensor is output. A physical information constrained neural network is constructed. A high-dimensional fusion feature tensor is used as input, and electromagnetic and acoustic physical equations and auxiliary sensor data are used as constraint terms of the loss function to solve the problem. The output is the three-dimensional positioning coordinates of the tunneling target in the tunnel global coordinate system.

[0007] As a preferred embodiment of the tunneling target location method based on multimodal data surveying of complex loose bodies described in this invention, the multimodal sensing data and tunneling condition data include: during the tunneling machine's movement, using a ground-penetrating radar antenna array installed behind the cutterhead to emit high-frequency pulse electromagnetic waves and collect electromagnetic wave time series data from the ground-penetrating radar. ; Acoustic time-series data of seismic waves were acquired using three-dimensional high-frequency seismic sensors deployed on a tunnel boring machine. Simultaneously read information from auxiliary sensors at the front and periphery of the cutter head. Real-time operating parameters of the tunneling machine.

[0008] As a preferred embodiment of the tunneling target localization method based on multimodal data surveying of complex loose bodies described in this invention, the adaptive noise reduction processing includes: converting acoustic time series data... Combined into a multidimensional observation matrix Suppose the observed signal is formed by mixing unknown source signals through a linear mixing matrix. Using a fast independent component analysis algorithm, find a demixing matrix. This makes the signals of the separated components... Maximize the non-Gaussianity of each other; Introducing the real-time operating parameters of the equipment, in the separated independent components In the process, the peak values ​​of the spectrum for each component are calculated; if the main peak of the spectrum highly coincides with the known mechanical operating frequency of the equipment, it is determined to be mechanical background noise and is removed; using the remaining independent components containing geological reflection information, the pure acoustic reflection signal is reconstructed from the inverse matrix of the unmixing matrix. ; Electromagnetic wave time series data Input a variational mode decomposition model, solve the constrained variational problem, and adaptively decompose it into... An eigenmode function with a specific center frequency; Perform a Hilbert transform on each modal component to obtain the instantaneous frequency; set a critical scattering frequency threshold corresponding to the particle size of the loose material. ; Center frequency greater than The weights of the modal components are set to zero, and the low-frequency modal components are retained for linear superposition and reconstruction to obtain a low-frequency radar signal that retains the characteristics of large-scale anomalies. As a preferred embodiment of the tunneling target localization method based on multimodal data surveying of complex loose bodies described in this invention, the depth mapping alignment includes processing the denoised low-frequency radar signal. With pure acoustic signal Perform Hilbert transform to construct an analytic signal; Calculate the modulus of the analytic signal to obtain the low-frequency radar energy envelope sequence. Acoustic energy envelope sequence The sequence highlights the energy extrema of the target body's reflective interface; Automatic alignment of radar envelope sequences without prior knowledge is performed using the DTW algorithm. Length is Acoustic envelope sequence Length is Calculate the Euclidean distance between points in the two sequences and construct... The local cost matrix; The minimum twist path is found based on the state transition equation. During the solution process, the path is forced to satisfy the condition that the endpoints coincide and that the path is strictly monotonically increasing. Based on the obtained optimal path, different time dimensions are... and Nonlinear resampling is performed, mapped to a dimensionless normalized relative depth space, and aligned to data features with the same index length. and .

[0009] As a preferred embodiment of the tunneling target localization method based on complex loose body multimodal data surveying described in this invention, the introduction of semantic prior vectors as context constraints includes dividing the registered relative depth space into a uniform voxel grid and calculating the gradient of the change in radar dielectric constant and acoustic impedance within each voxel. Set a baseline for background stratigraphic features. When the feature gradient within a voxel exceeds a set anomaly threshold... At that time, the current voxel is extracted as an entity node in a knowledge graph. ; For each extracted node Mounting high-dimensional attribute vectors This includes the spatial relative coordinates of the nodes, dielectric constant, acoustic impedance, and micro-Doppler frequency shift characteristics; Establish connections between nodes based on a pre-defined physics mechanism library. To form a three-dimensional spatial attribute knowledge graph Associated edges This includes spatial proximity edges, penetration mechanism edges, and acoustic coupling edges; A density-based spatial clustering algorithm is used to cluster the knowledge graph. The connected regions with dense weights and edge relationships are extracted to form several candidate anomaly subgraphs. ; Design drawing - text conversion serialization template, sub-drawing The overall topological form is transformed into natural language; the large model performs logical reasoning based on the built-in geological rescue knowledge chain, and simultaneously outputs confidence scores for each category, which are then normalized to generate a semantic category probability matrix. , as a semantic prior logical constraint.

[0010] As a preferred embodiment of the tunneling target localization method based on multimodal data surveying of complex loose bodies according to the present invention, wherein: the adaptive allocation of fusion weights for multimodal data features through a cross-attention mechanism includes, after alignment and High-dimensional physical feature sequences are extracted from the input bi-branch 1D-ResNet. and ; Using semantically guided bidirectional cross-attention, the generated semantic category probability matrix As a global contextual conditional attention injection mechanism; When the large model predicts that the area ahead is a water-rich region, the semantic category probability matrix The weights of easily attenuated radar features are forcibly reduced, while the weights of acoustic features are increased; the interaction features are calculated. and ; Adaptive gating weights are introduced to perform residual connections between the original features and the interaction features. Finally, the channels are concatenated and dimensionality reduced to output a high-dimensional fusion feature tensor containing all media properties of the target. ; As a preferred embodiment of the tunneling target localization method based on multimodal data surveying of complex loose bodies described in this invention, the step of solving the electromagnetic and acoustic physical equations and auxiliary sensor data as constraints of the loss function includes constructing a high-dimensional fusion feature tensor. and spatial coordinates The input solution network outputs the relative permittivity. Longitudinal wave velocity and target probability thermal value Using knowledge graphs to set and The upper and lower boundaries; A joint loss function is constructed, and the residuals of Maxwell's electromagnetic equations and the acoustic wave equations are used as physical evolution penalty terms by employing automatic differentiation techniques.

[0011] As a preferred embodiment of the tunneling target localization method based on multimodal data surveying of complex loose bodies described in this invention, wherein: the output of the three-dimensional positioning coordinates of the tunneling target in the tunnel global coordinate system includes, using auxiliary sensor data By combining the moisture content mapping and porosity mapping equations, macroscopic constraints for the sensor are constructed; and the semantic category probability matrix output by the large model is introduced. Apply KL divergence constraints to the network prediction results; After optimization and convergence, extract The three-dimensional probability extreme center, combined with the current global mileage coordinates of the tunneling machine and the cutterhead attitude angle, is transformed into absolute three-dimensional coordinates in the global tunnel coordinate system through a homogeneous transformation matrix.

[0012] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the tunneling target location method based on complex loose body multimodal data survey as described in the first aspect of the present invention.

[0013] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the tunneling target location method based on complex loose body multimodal data survey as described in the first aspect of the present invention.

[0014] The beneficial effects of this invention are as follows: It employs the Dynamic Time Warping (DTW) algorithm to construct a dimensionless unified relative depth space, completely eliminating time-depth conversion errors caused by the variable porosity and water content of complex loose bodies. It introduces semantic priors from the Large Geological Model (LLM) as global constraints, combined with a cross-attention mechanism. It senses local signal-to-noise ratios (such as radar attenuation in water-rich areas) and dynamically and intelligently allocates the fusion weights of electromagnetic and acoustic features, avoiding the contamination of overall data by invalid or high-noise modes. It constructs a Physical Information Constrained Neural Network (PINN), using the underlying electromagnetic and acoustic partial differential equations and multi-source in-situ sensor data (water pressure, stress) as joint penalty terms. Through dual locking of physical laws and sensor feedback, the inversion solution space is greatly narrowed, ensuring that the final output three-dimensional absolute coordinates have extremely high uniqueness and physical reliability, providing a guarantee for disaster prevention and mitigation across the entire tunnel cross-section. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a tunneling target location method based on multimodal data survey of complex loose bodies.

[0017] Figure 2 This is a schematic diagram of a computer device used for a tunneling target location method based on multimodal data surveying of complex loose bodies. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] Reference Figures 1-2 This is one embodiment of the present invention, which provides a method for locating tunneling targets based on multimodal data survey of complex loose bodies, including the following steps: S1: Synchronously collect multimodal sensor data and tunneling condition data during the tunneling process, and perform adaptive noise reduction processing.

[0022] Furthermore, the multimodal sensing data and tunneling condition data include, during the tunneling machine's movement, the use of a ground-penetrating radar antenna array mounted behind the cutterhead to emit high-frequency pulse electromagnetic waves and collect electromagnetic wave time-series data from the ground-penetrating radar. .

[0023] Acoustic time-series data of seismic waves were acquired using three-dimensional high-frequency seismic sensors deployed on a tunnel boring machine. Simultaneously read information from auxiliary sensors at the front and periphery of the cutter head. Real-time operating parameters of the tunneling machine.

[0024] An array-type ground-penetrating radar is employed, positioned behind the cutterhead or within the shield. The center frequency of the transmitting antenna is set to... (Typically in the hundreds of megahertz range, such as) To balance penetration depth and resolution. (Within the time window) Within, obtain the electromagnetic wave reflection time series data matrix. : in, This represents the number of receiving antenna channels. Indicates the first Each receiving antenna channel in time The instantaneous amplitude of the radar reflected echo signal recorded.

[0025] The acoustic response of the strata was acquired using three-dimensional high-frequency seismographs deployed on the tunnel sidewalls. The sampling rate was set to... (generally (), to obtain seismic wave time series data : in, The number of detector channels (including the three orthogonal components x, y, and z). Indicates the first Each detector channel in time The instantaneous amplitude of the recorded acoustic vibration signal.

[0026] Equipment operating condition and auxiliary sensor data acquisition and Synchronously acquire the real-time operating condition tensor of the TBM fed back by the main control PLC system. Compared with auxiliary sensor observations : in, For total thrust, The torque of the cutter head. The current three-dimensional mileage coordinates; The reading is from the hydraulic gauge at the front of the cutterhead. This represents the measured value of the surrounding rock stress. Indicates the ambient temperature sensor in time The real-time temperature measurement value.

[0027] Acoustic signal blind source noise reduction based on independent component analysis (ICA) is necessary because the noise energy generated during tunneling, such as the cutterhead breaking rock (continuous random vibration) and the motor operation (periodic electromagnetic interference), is much greater than the reflected echo energy from the strata ahead. Perform blind source separation.

[0028] The adaptive noise reduction process includes processing the acoustic time-series data. Combine them into a multidimensional observation matrix; suppose the observation signal is formed by mixing unknown source signals through a linear mixing matrix, and use a fast independent component analysis algorithm to find a demixing matrix that maximizes the non-Gaussianity of the separated component signals.

[0029] Assuming the observed seismic wave signal It is by A series of independent, unknown source signals Linearly mixed, firstly... Remove the mean, then calculate its covariance matrix and perform eigenvalue decomposition to obtain the whitening matrix. The converted data The components are uncorrelated and have a variance of 1.

[0030] Find a demixing matrix This makes the reconstructed signal Maximize the non-Gaussianity (mutual independence) of each component. Define negative entropy as a measure of non-Gaussianity.

[0031] By incorporating the equipment's real-time operating parameters, the peak spectral value of each component is calculated from the separated independent components. If the main peak of the spectrum highly overlaps with the equipment's known mechanical operating frequency, it is identified as mechanical background noise and eliminated. Using the remaining independent components containing geological reflection information, the pure acoustic reflection signal is reconstructed from the inverse matrix of the demixing matrix. .

[0032] Electromagnetic wave time series data Input a variational mode decomposition model, solve the constrained variational problem, and adaptively decompose it into... An eigenmode function with a specific center frequency.

[0033] Perform a Hilbert transform on each modal component to obtain the instantaneous frequency; set a critical scattering frequency threshold corresponding to the particle size of the loose material. .

[0034] Center frequency greater than The weights of the modal components are set to zero, and the low-frequency modal components are retained for linear superposition and reconstruction to obtain a low-frequency radar signal that retains the characteristics of large-scale anomalies. .

[0035] Let the discrete-time series of the ground-penetrating radar GPR signal after noise reduction be... , length is The discrete-time sequence of the seismic wave TSP signal is as follows: , length is .

[0036] S2: Extract the energy envelope feature sequence of multimodal sensing data, solve the optimal matching path through the dynamic time warping algorithm, and perform depth mapping alignment.

[0037] The depth mapping alignment includes processing the denoised low-frequency radar signal. With pure acoustic signal Perform a Hilbert transform to construct an analytic signal.

[0038] Feature envelope extraction and normalization based on Hilbert transform are necessary because radar electromagnetic waves and TSP acoustic waves differ greatly in amplitude, frequency, and physical dimensions, leading to significant errors when directly matching the original waveforms. Therefore, it is first necessary to extract dimensionless energy envelope features.

[0039] radar signals Harmony sound signal Calculate the analytic signals for each signal and extract the instantaneous amplitude envelope. For the signal... and Its envelope and The calculation formula is: in, This represents the Hilbert transform operator.

[0040] To eliminate the exponential energy decay difference caused by the probe depth, the extracted envelope sequence is subjected to Min-Max normalization and mapped to... The interval is used to obtain the normalized envelope sequence. and .

[0041] Calculate the modulus of the analytic signal to obtain the low-frequency radar energy envelope sequence. Acoustic energy envelope sequence The sequence highlights the energy extrema of the target body's reflective interface.

[0042] Automatic alignment of radar envelope sequences without prior knowledge is performed using the DTW algorithm. Length is Acoustic envelope sequence Length is Calculate the Euclidean distance between points in the two sequences and construct... The local cost matrix.

[0043] The minimum twist path is found based on the state transition equation. During the solution process, the path is forced to satisfy the condition that the endpoints coincide and that the path is strictly monotonically increasing.

[0044] Finding radar envelope sequences using the DTW algorithm With acoustic envelope sequence The non-linear alignment relationship between them.

[0045] calculate and Construct a Euclidean distance between all points in the middle. Distance matrix The first in the matrix line, number The elements of a column are defined as follows: Constructing the cumulative cost matrix To find the alignment path that minimizes the total distance, a state transition equation is defined. For and : Boundary conditions are initialized as follows: , , .

[0046] Backtracking optimization generates a twisted path: starting from the end of the cumulative cost matrix Start by searching backwards for the predecessor node that minimizes the cumulative distance, until you return to the starting point. This yields an optimal alignment path. in, Indicates the first At the matching step, the radar signal's first... The sampling point and the first sound wave signal Each sampling point corresponds to the same physical reflection interface in geological space, and the path length satisfies: Using the obtained optimal matching path This maps heterogeneous signals in the time domain to a unified spatial domain.

[0047] Establish a dimensionless normalized relative depth space. The sequence of steps to match the path. As a baseline, it is mapped proportionally to axis: According to the path The original radar signal value and the value of the original acoustic signal Assign this depth node For one-to-many or many-to-one mapping points (i.e., regions where the signal is "stretched" or "compressed"), cubic spline interpolation is used for smoothing. Finally, the output is processed at the same relative depth coordinates. Strictly aligned radar feature vectors Harmony sound wave eigenvectors This serves as the direct input for the next step of the cross-modal attention network.

[0048] Based on the obtained optimal path, different time dimensions are... and Nonlinear resampling is performed, mapped to a dimensionless normalized relative depth space, and aligned to data features with the same index length. and .

[0049] S3: Input the aligned multimodal data features into the dual-branch feature encoding network, introduce semantic prior vectors as context constraints, adaptively allocate the fusion weights of the multimodal data features through the cross-attention mechanism, and output a high-dimensional fusion feature tensor.

[0050] The introduction of semantic prior vectors as contextual constraints includes dividing the registered relative depth space into a uniform voxel grid and calculating the gradient of the change in radar dielectric constant and acoustic impedance within each voxel.

[0051] Set a baseline for background stratigraphic features. When the feature gradient within a voxel exceeds a set anomaly threshold... At that time, the current voxel is extracted as an entity node in a knowledge graph. .

[0052] Based on feature gradient-based spatial voxelization and knowledge graph node extraction, the registered 3D normalized relative depth space is divided into... A uniform voxel grid. Let any voxel be... .

[0053] Calculate each voxel Internal radar dielectric constant and acoustic impedance The spatial gradient of change, i.e. and .

[0054] Based on historical normal exploration data, a baseline for background stratigraphic characteristics and corresponding anomaly gradient thresholds are established. and If the current voxel satisfies or .

[0055] The anomalous voxel is then extracted into a three-dimensional spatial attribute knowledge graph. One of the entity nodes (EntityNode) For each extracted node Mount a high-dimensional attribute feature vector : in, Let be the spatial relative coordinates of the node in the voxel mesh. Where is the dielectric constant. For acoustic impedance, The extracted micro-Doppler frequency shift features.

[0056] For each extracted node Mounting high-dimensional attribute vectors This includes the spatial relative coordinates of the nodes, dielectric constant, acoustic impedance, and microDoppler frequency shift characteristics.

[0057] Establish connections between nodes based on a pre-defined physics mechanism library. To form a three-dimensional spatial attribute knowledge graph Associated edges This includes spatial proximity edges, permeation mechanism edges, and acoustic coupling edges.

[0058] A density-based spatial clustering algorithm is used to cluster the knowledge graph. The connected regions with dense weights and edge relationships are extracted to form several candidate anomaly subgraphs. .

[0059] Design drawing - text conversion serialization template, sub-drawing The overall topological form is transformed into natural language; the large model performs logical reasoning based on the built-in geological rescue knowledge chain, and simultaneously outputs confidence scores for each category, which are then normalized to generate a semantic category probability matrix. , as a semantic prior logical constraint.

[0060] The adaptive allocation of fusion weights for multimodal data features via cross-attention mechanism includes aligning the weights... and High-dimensional physical feature sequences are extracted from the input bi-branch 1D-ResNet. and .

[0061] Using semantically guided bidirectional cross-attention, the generated semantic category probability matrix As a global contextual conditional attention mechanism.

[0062] When the large model predicts that the area ahead is a water-rich region, the semantic category probability matrix The weights of easily attenuated radar features are forcibly reduced, while the weights of acoustic features are increased; the interaction features are calculated. and .

[0063] Adaptive gating weights are introduced to perform residual connections between the original features and the interaction features. Finally, the channels are concatenated and dimensionality reduced to output a high-dimensional fusion feature tensor containing all media properties of the target. .

[0064] Iterate through all extracted node sets. Establish a set of associated edges between nodes based on a pre-defined physical mechanism library. Associative edges include three types: Spatial proximity edge: Calculate the Euclidean distance between nodes, and establish an edge between nodes whose distance is less than a set threshold.

[0065] Permeation mechanism edge: Determine the pore connectivity based on the direction of the dielectric constant gradient and establish medium permeation connection edges.

[0066] Acoustic coupling edge: an acoustic reflection and transmission coupling edge established based on the difference in acoustic impedance.

[0067] Density-based spatial clustering algorithms (such as DBSCAN) are used to cluster knowledge graphs. Clustering is performed on connected regions with dense node distribution and high edge association weights (i.e., physical attribute similarity) to form... Candidate anomaly subgraphs.

[0068] Design a standardized Prompt graph-to-text conversion serialization template to transform the topological structure of subgraphs into natural language sequences. (In relative coordinates...) There is a region with high dielectric constant and low acoustic impedance, and the spatially adjacent edges are densely connected with the permeation mechanism edges.

[0069] The above natural language sequence is input into the vertical geological large-scale model LLM. The large model calls the built-in geological rescue knowledge chain to perform logical reasoning and outputs a confidence score vector for the geological categories that may be encountered ahead (such as: water-rich faults, rock clusters, cavities, etc.). .

[0070] For the rating vector Perform Softmax normalization to generate a prior probability matrix that includes semantic categories. This serves as a global semantic context constraint for the next step of feature fusion.

[0071] The spatiotemporally registered radar sequence Harmony and acoustic sequences The parameters are respectively input into a two-branch one-dimensional residual convolutional neural network (1D-ResNet) with non-shared parameters for encoding.

[0072] After multiple convolutional and downsampling operations, a deep, high-dimensional physical feature sequence is extracted: in, For sequence length, This represents the number of feature channels. and This represents the mapping function of a two-branch one-dimensional residual convolutional neural network used to process radar and acoustic signals. This indicates the dimension of the real space to which the generated feature matrix belongs.

[0073] The semantic category probability matrix generated by the large model Mapped to the same dimension as the features, it is injected into the cross-attention module as a global context condition. Taking radar features guiding acoustic features as an example, the radar features are... After semantic constraint correction, it is mapped to a query term, which includes sound wave features. Mapped to keys and values: Calculate interaction features based on the attention score matrix: It should be noted that the characteristics of acoustic wave guided radar The calculation logic is symmetrical to this, and the result is obtained using the same formula as above. .

[0074] in, This represents the learnable linear mapping weight matrix of the query item. This represents the balance coefficient. This represents the semantically corrected radar query feature matrix. and These represent the linear mapping weight matrices for keys and values, respectively. This represents the matrix transpose operation. Indicates the scaling dimension.

[0075] If the large model predicts that the area ahead is a water-rich region... The injection will forcibly suppress the activation values ​​of radar features that are susceptible to moisture decay in the attention matrix through the Softmax function, while increasing the attention weight of acoustic features, thereby enabling the network to adaptively assign trust to multimodal features.

[0076] Adaptive gated residual connections and tensor outputs are introduced, employing a learnable gating mechanism to generate an adaptive weight matrix that controls the fusion ratio. and (The range is in) between).

[0077] We perform a weighted residual connection between the original features and the interactive features output by cross-attention, preserving the feature noise floor: in, This is for element-wise multiplication.

[0078] The two features after residual connection are concatenated along the channel dimension, and finally passed through a linear mapping layer (such as...). Convolutional dimensionality reduction and integration outputs a high-dimensional fused feature tensor containing comprehensive medium properties of the target. This tensor will serve as the direct input to the downstream inversion network.

[0079] S4: Construct a physical information constrained neural network, take a high-dimensional fusion feature tensor as input, use electromagnetic and acoustic physical equations and auxiliary sensor data as constraint terms of the loss function to solve, and output the three-dimensional positioning coordinates of the tunneling target in the tunnel global coordinate system.

[0080] The method of solving the electromagnetic and acoustic physical equations and auxiliary sensor data as constraints of the loss function includes constructing a system based on... and spatial coordinates The input solution network outputs the relative permittivity. Longitudinal wave velocity and target probability thermal value Using knowledge graphs to set and The upper and lower boundaries.

[0081] A fully connected multilayer perceptron (MLP) is constructed as the inversion solution network. The high-dimensional fusion feature tensor is then used. As a conditional input to the hidden layer of the network, the three-dimensional relative spatial coordinates As input variables, the network outputs three core parameters corresponding to this coordinate point in parallel: in, For network learnable parameters, The relative permittivity, For longitudinal wave velocity, The probability heat value represents the probability that an anomalous target (such as a boulder or cavity) exists at that coordinate point, as output by the network. This represents the inversion computation network.

[0082] To prevent the network from getting trapped in non-physical local minima during the early training phase, entity node attributes extracted from the knowledge graph in previous steps are invoked. Based on the geological benchmark of the current tunneling mileage, physical upper and lower bounds for dielectric constant and wave velocity are extracted. A scaled sigmoid activation function or boundary truncation function is used to strongly constrain the network's original output layer, forcing the network's output medium parameters to fall within the physically permissible range.

[0083] A joint loss function is constructed, and the residuals of Maxwell's electromagnetic equations and the acoustic wave equations are used as physical evolution penalty terms by employing automatic differentiation techniques.

[0084] Using the automatic differentiation technique built into deep learning frameworks, the network output is calculated with respect to the input space coordinates. and time The partial derivatives of each order.

[0085] Electromagnetic evolution penalty term, using Maxwell's electromagnetic equations as residual terms. Constraining radar feature propagation paths: The sound wave evolution penalty term treats the sound wave equation as a residual term. Constraining the characteristic propagation path of sound waves: Extract the PDE loss function, randomly sample points in the probe space domain, and calculate the mean square error as a physical evolution penalty term. : in, It is the Laplace operator, representing the sum of the second-order partial derivatives with respect to spatial coordinates. Indicates time Inter-variable The second-order partial derivatives of . Represents spatial coordinates and time The electromagnetic wave field distribution value at that location. Representing a spatial point The relative permittivity of the medium at that location. It represents the speed of light in a vacuum. Represents spatial coordinates and time The acoustic field distribution value at the location (representing the seismic wave displacement field or stress field reconstructed by the network).

[0086] Representing a spatial point The longitudinal wave velocity of the strata at that location is directly related to the density and stress state of the rock mass.

[0087] The output three-dimensional positioning coordinates of the tunneling target in the tunnel global coordinate system include those obtained using auxiliary sensor data. By combining the moisture content mapping and porosity mapping equations, macroscopic constraints for the sensor are constructed; and the semantic category probability matrix output by the large model is introduced. KL divergence constraints are applied to the network prediction results.

[0088] To eliminate the ambiguity caused by isospectral heterogeneity (for example, high water content and high density rock masses may produce similar echo characteristics), external macroscopic data are incorporated into the loss function.

[0089] Auxiliary sensor data collected by devices such as cutterhead hydraulic gauges and stress gauges are introduced. A moisture content mapping equation is constructed using the Topp empirical formula. A porosity mapping equation was constructed using the Wyllie time-averaged equation. The calculated dielectric constant / wave velocity conversion values ​​predicted by the network are compared with actual sensor observations (such as observed water pressure). and stress The residuals between ) .

[0090] Introducing the semantic category probability matrix output by the logical reasoning of the large model .because It includes macroscopic geological common sense judgment, which is then compared with the target probability distribution. Calculate the Kulbeck-Leibler divergence and apply a penalty: Construct the final global joint loss function The formula is expressed as: in, The waveform data fitting error is the difference between the waveform reconstructed by the network and the actual received waveform. These are the balancing weight coefficients dynamically adjusted using a gradient normalization algorithm. The gradient normalization algorithm... Automatic adjustments are made in each iteration to ensure a smooth progression of multi-task learning.

[0091] Update network parameters using backpropagation with the Adam optimizer This continues until the joint loss function converges, at which point the network outputs a probability heat value that has a unique optimal solution.

[0092] After optimization and convergence, the continuous output of the network is... A gridded evaluation is performed, a threshold is set to extract high-probability connected components, and the three-dimensional probability extremum center of the connected component is calculated to obtain the relative spatial coordinates of the target in the local coordinate system of the tunneling machine. Obtain the absolute mileage coordinates of the tunnel boring machine in the tunnel's global coordinate system. And the orientation angle of the tool holder. Construct the spatial homogeneous transformation matrix. This maps local coordinates to global coordinates.

[0093] After optimization and convergence, extract The three-dimensional probability extreme center, combined with the current global mileage coordinates of the tunnel boring machine and the cutterhead attitude angle, is transformed into absolute three-dimensional coordinates in the global tunnel coordinate system through a homogeneous transformation matrix.

[0094] This embodiment also provides a computer device, such as... Figure 2 As shown, the method for locating tunneling targets based on multimodal data surveying of complex loose bodies includes: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the tunneling target location method based on multimodal data surveying of complex loose bodies as proposed in the above embodiments.

[0095] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0096] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the tunneling target location method based on complex loose body multimodal data surveying as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0097] In summary, this invention achieves the following: Simultaneous acquisition of multimodal sensor data and tunneling condition data during the tunneling process, followed by adaptive noise reduction; extraction of energy envelope feature sequences from the multimodal sensor data, calculation of the optimal matching path using a dynamic time warping algorithm, and depth mapping alignment; input of the aligned multimodal data features into a dual-branch feature encoding network, introduction of semantic prior vectors as contextual constraints, adaptive allocation of fusion weights for the multimodal data features through a cross-attention mechanism, and output of a high-dimensional fusion feature tensor; and construction of a physical information constraint neural network, using the high-dimensional fusion feature tensor as input, and solving for electromagnetic and acoustic physical equations and auxiliary sensor data as constraints in the loss function, outputting the three-dimensional positioning coordinates of the tunneling target in the tunnel's global coordinate system.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for locating tunneling targets based on multimodal data surveying of complex loose bodies, characterized in that: This includes synchronously acquiring multimodal sensor data and tunneling condition data during the tunneling process, and performing adaptive noise reduction processing; Extract the energy envelope feature sequence from the multimodal sensing data, solve the optimal matching path using the dynamic time warping algorithm, and perform depth mapping alignment; The aligned multimodal data features are input into a dual-branch feature encoding network. A semantic prior vector is introduced as a context constraint. The fusion weights of the multimodal data features are adaptively allocated through a cross-attention mechanism, and a high-dimensional fusion feature tensor is output. A physical information constrained neural network is constructed. A high-dimensional fusion feature tensor is used as input, and electromagnetic and acoustic physical equations and auxiliary sensor data are used as constraint terms of the loss function to solve the problem. The output is the three-dimensional positioning coordinates of the tunneling target in the tunnel global coordinate system.

2. The tunneling target localization method based on multimodal data survey of complex loose bodies as described in claim 1, characterized in that: The multimodal sensing data and tunneling condition data include, during the tunneling machine's movement, the use of a ground-penetrating radar antenna array mounted behind the cutterhead to emit high-frequency pulse electromagnetic waves and collect time-series electromagnetic wave data from the ground-penetrating radar. ; Acoustic time-series data of seismic waves were acquired using three-dimensional high-frequency seismic sensors deployed on a tunnel boring machine. ; Synchronously read data from auxiliary sensors at the front and around the cutter head. Real-time operating parameters of the tunneling machine.

3. The tunneling target localization method based on multimodal data survey of complex loose bodies as described in claim 2, characterized in that: The adaptive noise reduction process includes processing the acoustic time-series data. Combined into a multidimensional observation matrix; Suppose the observed signal is formed by mixing unknown source signals through a linear mixing matrix. Using the fast independent component analysis algorithm, we can find a demixing matrix that maximizes the non-Gaussianity among the separated component signals. By incorporating the equipment's real-time operating parameters, the peak spectral value of each component is calculated from the separated independent components. If the main peak of the spectrum highly overlaps with the equipment's known mechanical operating frequency, it is identified as mechanical background noise and eliminated. Using the remaining independent components containing geological reflection information, the clean acoustic signal is reconstructed from the inverse matrix of the demixing matrix. ; Electromagnetic wave time series data Input a variational mode decomposition model, solve the constrained variational problem, and adaptively decompose it into... An eigenmode function with a specific center frequency; Perform a Hilbert transform on each modal component to obtain the instantaneous frequency; set a critical scattering frequency threshold corresponding to the particle size of the loose material. ; Center frequency greater than The weights of the modal components are set to zero, and the low-frequency modal components are retained for linear superposition and reconstruction to obtain a low-frequency radar signal that retains the characteristics of large-scale anomalies. .

4. The tunneling target localization method based on multimodal data survey of complex loose bodies as described in claim 3, characterized in that: The depth mapping alignment includes processing the denoised low-frequency radar signal. With pure acoustic signal Perform Hilbert transform to construct an analytic signal; Calculate the modulus of the analytic signal to obtain the low-frequency radar energy envelope sequence. Acoustic energy envelope sequence The sequence highlights the energy extrema of the target body's reflective interface; The DTW algorithm is used for automatic alignment of wave velocities without prior knowledge. The low-frequency radar energy envelope sequence is assumed to be... Length is Acoustic energy envelope sequence Length is Calculate the Euclidean distance between points in the two sequences and construct... The local cost matrix; The minimum twist path is found based on the state transition equation. During the solution process, the path is forced to satisfy the condition that the endpoints coincide and that the path is strictly monotonically increasing. Based on the obtained optimal path, different time dimensions are... and Nonlinear resampling is performed, mapped to a dimensionless normalized relative depth space, and aligned to data features with the same index length. and ,in, Represents the radar feature vector. This represents the characteristic vector of the sound wave.

5. The tunneling target localization method based on multimodal data survey of complex loose bodies as described in claim 4, characterized in that: The introduction of semantic prior vectors as context constraints includes dividing the registered relative depth space into a uniform voxel grid and calculating the gradient of the change in radar dielectric constant and acoustic impedance within each voxel. Set a baseline for background stratigraphic features. When the feature gradient within a voxel exceeds a set anomaly threshold... At that time, the current voxel is extracted as an entity node in a knowledge graph. ; For each extracted node Mounting high-dimensional attribute vectors This includes the spatial relative coordinates of the nodes, dielectric constant, acoustic impedance, and micro-Doppler frequency shift characteristics; Establish connections between nodes based on a pre-defined physics mechanism library. To form a three-dimensional spatial attribute knowledge graph Associated edges This includes spatial proximity edges, penetration mechanism edges, and acoustic coupling edges; A density-based spatial clustering algorithm is used to cluster the knowledge graph. The connected regions with dense weights and edge relationships are extracted to form several candidate anomaly subgraphs; The design diagram-text conversion serialization template transforms the overall topological form of the subgraphs into natural language; the large model performs logical reasoning based on the built-in geological rescue knowledge chain, simultaneously outputting confidence scores for each category, and generating a semantic category probability matrix after normalization. , as a semantic prior logical constraint.

6. The tunneling target localization method based on multimodal data surveying of complex loose bodies as described in claim 5, characterized in that: The adaptive allocation of fusion weights for multimodal data features via cross-attention mechanism includes aligning the weights... and High-dimensional physical feature sequences are extracted from the input bi-branch 1D-ResNet. and ,in, Indicates radar characteristics, Indicates the characteristics of sound waves; Using semantically guided bidirectional cross-attention, the generated semantic category probability matrix As a global contextual conditional attention injection mechanism; When the large model predicts that the area ahead is a water-rich region, the semantic category probability matrix The weights of easily attenuated radar features are forcibly reduced, while the weights of acoustic features are increased; the interaction features are calculated. and , This indicates the characteristics of acoustically guided radar. Adaptive gating weights are introduced to perform residual connections between the original features and the interaction features. Finally, the channels are concatenated and dimensionality reduced to output a high-dimensional fusion feature tensor containing all media properties of the target. .

7. The tunneling target localization method based on multimodal data survey of complex loose bodies as described in claim 6, characterized in that: The step of solving the electromagnetic and acoustic physical equations and auxiliary sensor data as constraints of the loss function includes constructing a high-dimensional fused feature tensor. and spatial coordinates The input solution network outputs the relative permittivity. Longitudinal wave velocity and target probability thermal value Using knowledge graphs to set and The upper and lower boundaries; A joint loss function is constructed, and the residuals of Maxwell's electromagnetic equations and the acoustic wave equations are used as physical evolution penalty terms by employing automatic differentiation techniques.

8. The tunneling target localization method based on multimodal data survey of complex loose bodies as described in claim 7, characterized in that: The output three-dimensional positioning coordinates of the tunneling target in the tunnel global coordinate system include those obtained using auxiliary sensor data. By combining the moisture content mapping and porosity mapping equations, macroscopic constraints for the sensor are constructed; and the semantic category probability matrix output by the large model is introduced. Apply KL divergence constraints to the network prediction results; After optimization and convergence, extract The three-dimensional probability extreme center, combined with the current global mileage coordinates of the tunneling machine and the cutterhead attitude angle, is transformed into absolute three-dimensional coordinates in the global tunnel coordinate system through a homogeneous transformation matrix.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the tunneling target location method based on multimodal data survey of complex loose bodies as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the tunneling target location method based on multimodal data survey of complex loose bodies as described in any one of claims 1 to 8.

Images