Geological structure intelligent detection system and method for deep special space

Abstract

The invention provides a geological structure intelligent detection system and method for a deep special space, and relates to the technical field of geological exploration. The geological structure intelligent detection system and method for the deep special space comprises the steps that S1, a distributed sensor network is deployed in the deep special space, and seismic wave, electromagnetic field, sound wave, temperature and strain multi-physical field original detection data are synchronously collected; constructing a noise channel model based on a quantum computing principle, and modeling the original detection data; a quantum inspired adaptive filtering algorithm is applied, and noise characteristics are estimated in real time through variational optimization. Through synchronous acquisition of a distributed multi-physics field sensor network and an adaptive filtering technology of quantum inspiration, effective enhancement of weak characteristic signals is realized in a strong noise environment, and the signal-to-noise ratio and the target geological structure identification precision are significantly improved; and a neural network inversion model is combined with physical information constraint.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, specifically to an intelligent geological structure detection system and method for deep, special spaces. Background Technology

[0002] The field of geological exploration technology involves the detection and analysis of underground rock strata structures, mineral resource distribution, and geological engineering conditions, with a focus on solving the problems of data acquisition, interpretation, and modeling in complex geological environments. This field encompasses geophysical exploration (seismic waves, electromagnetic fields, gravity fields, etc.), geotechnical parameter inversion, and geological hazard risk assessment. Particularly for special spatial environments such as deep high ground stress, high temperature and pressure, and strong electromagnetic interference, it is necessary to overcome the signal-to-noise ratio limitations, multiple solutions dilemmas, and real-time bottlenecks of traditional detection methods. Its technological development relies on the deep integration of multi-source sensor networks, artificial intelligence algorithms, and numerical simulation, aiming to provide accurate geological data for mineral resource development, tunnel engineering safety, and geological disaster prevention. One such intelligent geological structure detection system and method for deep, special spaces refers to: achieving data enhancement under strong interference environments through distributed multi-physics field sensor networks and quantum-inspired noise reduction technology; constructing a high-precision three-dimensional geological parameter model by combining a physically constrained inversion model and a federated learning framework; and integrating knowledge-guided interpretability analysis and multi-field coupled digital twin technology to form a closed-loop intelligent detection system from data acquisition to dynamic risk projection.

[0003] Traditional geological exploration methods face significant limitations in deep and unique spaces: geophysical exploration is affected by strong electromagnetic interference and high ground stress, resulting in a severely degraded signal-to-noise ratio of the original signal, causing weak geological features to be submerged by noise, such as a significantly high rate of missed detection of microseismic events in mines; the inversion of geotechnical parameters relies on single physical field data, lacking multimodal feature fusion and physical law constraints, leading to large errors in fault boundary interpretation and inducing water inrush accidents at tunnel faces; distributed sensor nodes adopt a centralized data processing architecture, and the edge end cannot interpret data in real time, resulting in significant data transmission delays and missed rockburst precursor identification windows; risk assessment models ignore prior geological knowledge and uncertainty quantification, and the interpretation conclusions are not traceable, causing deviations in engineering support scheme design; numerical simulation systems operate statically and do not integrate real-time sensor data update mechanisms, making it impossible to predict sudden changes in the seepage field under excavation disturbances. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent geological structure detection system and method for deep and special spaces. It solves the significant limitations of traditional geological exploration methods in deep and special spaces: geophysical exploration is affected by strong electromagnetic interference and high ground stress, resulting in a severe deterioration of the original signal-to-noise ratio, which causes weak geological features to be submerged by noise, such as a significantly high rate of missed detection of microseismic events in mines; the inversion of geotechnical parameters relies on single physical field data and lacks multimodal feature fusion and physical law constraints, resulting in large errors in fault boundary interpretation and inducing water inrush accidents at tunnel faces.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent geological structure detection system and method for deep, special spaces, comprising the following steps:

[0006] S1: Deploy a distributed sensor network in a deep, special space to simultaneously collect raw detection data of multiple physical fields, including seismic waves, electromagnetic fields, sound waves, temperature, and strain; construct a noise channel model based on quantum computing principles to model the raw detection data; apply a quantum-inspired adaptive filtering algorithm to estimate noise characteristics in real time through variational optimization and dynamically adjust filtering parameters to process the raw detection data, thereby achieving strong background noise suppression and enhancement of weak feature signals related to the target geological structure, and generating a preprocessed signal dataset;

[0007] S2: Based on the preprocessed signal dataset, depth features of each physical field data are extracted using a convolutional neural network model; a feature fusion module based on an attention mechanism is used to dynamically weight and fuse the depth features to generate a joint feature representation; the joint feature representation is input into a neural network inversion model constrained by physical information, which uses the control equations describing the geophysical response as regularization terms embedded in the loss function to perform three-dimensional spatial geological physical parameter inversion and geological interface identification, generating a three-dimensional geological physical parameter model;

[0008] S3: Based on the background geological framework provided by the 3D geological physical parameter model and the real-time data stream in the preprocessed signal dataset; on the edge computing nodes deployed at the front end of the detection, a lightweight deep learning model is used to perform preliminary geological event detection and classification on the local real-time data; a federated learning architecture is constructed, where edge nodes only upload model updates to the central server, and the central server aggregates and updates to generate a global interpretation model and distributes it; edge nodes combine the local model, the global interpretation model, and the background geological framework to perform collaborative interpretation of the real-time data; based on the interpretation results and pre-set risk assessment rules, the geological hazard risk level is calculated in real time, generating real-time geological state interpretation data and risk level distribution data;

[0009] S4: Based on a three-dimensional geological physical parameter model, real-time geological state interpretation data and risk level distribution data, as well as a pre-set regional geological prior knowledge base; construct a knowledge-guided graph neural network model, and use geological structural framework constraint feature propagation and interpretation; apply interpretability analysis algorithms to analyze the decision basis of key models in S2 and S3; use uncertainty quantification technology to estimate the confidence level of key geological interface locations and risk level assessment results, and generate interpretable geological structure probability models and quantitative risk assessment data;

[0010] S5: Integrate a three-dimensional geological physical property parameter model, real-time geological state interpretation data and risk level distribution data, interpretable geological structure probability model and quantitative risk assessment data, and real-time multi-physics field data stream from the preprocessed signal dataset; establish a geological-stress-seepage multi-field coupled numerical simulation model; apply a data assimilation algorithm to dynamically invert and update the key time-varying parameters of the simulation model using the real-time multi-source data; construct a digital twin that integrates geological structure, real-time sensing state, and updated multi-field information; based on the digital twin, perform multi-physics field short-term evolution trend prediction and engineering disturbance scenario simulation and risk assessment, and generate a dynamic deep space geological-engineering multi-field coupled digital twin and prediction data.

[0011] Preferably, step S1 includes the following steps:

[0012] S101: Deploy a distributed sensor network in a special deep space to synchronously collect raw seismic wave, electromagnetic field, sound wave, temperature and strain data to generate raw multiphysics time series datasets;

[0013] S102: Based on the original multiphysics time series dataset, using the principle of quantum computing, the signal is regarded as a "quantum state" and the noise is simulated as a "quantum noise channel", a hybrid signal-noise mathematical model is constructed, and a quantum noise channel model is generated;

[0014] S103: Based on the quantum noise channel model, the variational optimization algorithm is applied to iteratively estimate the type, intensity, and correlation parameters of noise in real time, and generate a dynamic noise characteristic parameter set;

[0015] S104: Based on the original multiphysics time series dataset and dynamic noise characteristic parameter set, a quantum-inspired adaptive filtering algorithm is adopted to dynamically adjust the filter parameters, suppress background noise and enhance the weak geological features of the target, and generate a feature enhancement signal.

[0016] S105: Based on feature enhancement signals, timestamp synchronization and spatial interpolation algorithms are used to unify the spatiotemporal reference of different sensors, eliminate acquisition delay and spatial offset, and generate a preprocessed signal dataset.

[0017] Preferably, step S2 includes the following steps:

[0018] S201: Based on the preprocessed signal dataset, an improved convolutional neural network is used to extract high-dimensional abstract features from seismic, electromagnetic, and acoustic data respectively, generating a single-modal deep feature set;

[0019] S202: Based on a single-modal deep feature set, a gated attention mechanism is used to dynamically calculate the spatial weights of different modal features, perform cross-modal weighted fusion, and generate a multimodal joint feature tensor.

[0020] S203: Based on the multimodal joint feature tensor, a physical information constrained neural network is constructed. The elastic wave equation / Maxwell equation is embedded as a regularization term in the loss function, and the constrained inversion process conforms to physical laws, generating a physical regularized inversion model.

[0021] S204: Based on the physical regularization inversion model, the gradient descent optimization algorithm is used to iteratively update physical property parameters such as velocity, density, and resistivity until the loss function converges, generating a three-dimensional geological physical property parameter grid model.

[0022] Preferably, step S3 includes the following steps:

[0023] S301: Deploy lightweight convolutional networks with knowledge distillation and compression on edge devices such as mining robots and sensor nodes to support real-time data processing and generate edge intelligent interpretation models;

[0024] S302: Based on the preprocessed signal dataset and the edge intelligent interpretation model, microseismic event localization and electromagnetic anomaly classification are performed at the edge, generating a local geological event label set;

[0025] S303: Based on a local geological event tag set, a federated learning framework is adopted, with edge nodes uploading encrypted model gradients to the central server; the server generates a global model through a secure aggregation algorithm, and generates a federated optimized global interpretation model;

[0026] S304: Based on a federated optimization global interpretation model, a three-dimensional geological property parameter model, and a local geological event tag set, data is fused at the edge to identify changes in rock mass stability and geological anomalies, and to generate a dynamic geological state interpretation map.

[0027] S305: Based on the dynamic geological state interpretation map, a random forest risk assessment model is adopted, combined with a pre-set rule base, to output spatial risk levels and generate real-time geological state interpretation and risk distribution data.

[0028] Preferably, step S4 includes the following steps:

[0029] S401: Integrating regional borehole data, structural maps, and lithological sequences, and using graph database technology, a geological prior knowledge graph containing stratigraphic contact relationships, fault attributes, and physical property ranges is constructed to generate a structured geological knowledge base;

[0030] S402: Based on a three-dimensional geological physical property parameter model, real-time geological state interpretation and risk distribution data, and a structured geological knowledge base, a graph neural network is constructed to generate knowledge-guided geological structure interpretation results by constraining feature propagation through knowledge graph node relationships.

[0031] S403: Based on the knowledge-guided geological structure interpretation results, the contribution of input features to fault identification decisions is quantified using the Shapley sum interpretation method or gradient salience map, and a geological interpretation attribution report is generated.

[0032] S404: Based on the knowledge-guided geological structure interpretation results, Monte Carlo Dropout technology is used to perform multiple sampling predictions on the locations of key geological interfaces, calculate confidence intervals, and generate geological structure confidence interval data.

[0033] S405: Integrates geological interpretation attribution reports, geological structure confidence interval data, real-time geological status interpretation and risk distribution data, generates a risk assessment with confidence level, and generates an interpretable geological structure probability model and quantitative risk assessment data.

[0034] Preferably, step S5 includes the following steps:

[0035] S501: Based on the three-dimensional geological physical property parameter model, establish the rock mass deformation-seepage control equation, initialize the stress field and seepage field parameters, and generate a multi-physics field coupled simulation model.

[0036] S502: Based on a multi-physics field coupled simulation model, preprocessed signal dataset, real-time geological state interpretation and risk distribution data, a multi-physics field coupled model is generated by dynamically inverting and updating permeability and rock mass modulus parameters using ensemble Kalman filtering.

[0037] S503: Based on a calibrated multi-field coupling model, an interpretable geological structure probability model, and quantitative risk assessment data, a three-dimensional mesh model containing parameter uncertainties is constructed to generate a probabilistic digital twin base.

[0038] S504: Based on a probabilistic digital twin base, using finite element numerical simulation, it calculates the stress field redistribution and seepage path expansion trend in the next 24 hours, and generates multiphysics short-term prediction data.

[0039] S505: Based on a probabilistic digital twin base, it injects engineering disturbance parameters, simulates the evolution of surrounding rock damage and water inrush risk, outputs risk warning maps, and generates dynamic geological-engineering multi-field coupled digital twins and prediction data.

[0040] A smart geological structure detection system for deep and special spaces includes the following modules: a multi-source sensing and processing module, which uses a quantum-inspired adaptive filtering algorithm to suppress noise in raw seismic wave, electromagnetic field, and acoustic wave data based on a deployed distributed sensor network; unifies the multi-source data benchmark through a spatiotemporal alignment algorithm; and outputs an enhanced standardized data stream to generate a preprocessed signal dataset.

[0041] The intelligent interpretation and assessment module, based on a preprocessed signal dataset, uses a physical information-constrained neural network to invert 3D geological parameters; it achieves edge-cloud collaborative interpretation through a federated learning framework; and it combines a knowledge-guided graph neural network to output interpretable risk assessments, generating 3D geological property parameter models and quantitative risk assessment data.

[0042] The twin decision support module constructs a rock mass deformation-seepage coupling model based on a three-dimensional geological physical parameter model and quantitative risk assessment data; it uses an ensemble Kalman filter algorithm to dynamically assimilate real-time data; and it generates engineering disturbance response predictions through a probabilistic twin engine, thus generating a dynamic multi-field coupled digital twin.

[0043] Preferably, the multi-source sensing processing module includes a heterogeneous acquisition submodule, a quantum noise reduction submodule, and a spatiotemporal alignment submodule;

[0044] The heterogeneous acquisition submodule deploys arrays of seismic, electromagnetic, and acoustic sensors to simultaneously acquire raw physical field data from deep space and generate raw multiphysics datasets.

[0045] The quantum noise reduction submodule, based on the original multiphysics dataset, constructs a hybrid model using a quantum noise channel modeling algorithm; dynamically estimates noise parameters using a variational optimization algorithm; and suppresses noise and generates a feature-enhanced signal set through a quantum-inspired adaptive filter.

[0046] The spatiotemporal alignment submodule, based on the feature-enhanced signal set, uses a timestamp synchronization algorithm to eliminate acquisition delay; it unifies the spatial reference through Kriging spatial interpolation and generates a preprocessed signal dataset.

[0047] Preferably, the intelligent interpretation and evaluation module includes a property inversion submodule, a federated interpretation submodule, and a knowledge analysis submodule;

[0048] The physical property inversion submodule, based on a preprocessed signal dataset, uses a deep residual convolutional network to extract single-modal features; it achieves cross-modal fusion through a gated attention mechanism; it constructs a physically constrained inversion model to output three-dimensional physical property parameters and generates a three-dimensional geological physical property mesh model.

[0049] The federated interpretation submodule is based on a 3D geological property mesh model and deploys a lightweight convolutional network at the edge. It uses a secure aggregation federated learning framework to update the global model and combines real-time data streams to identify geological anomalies and generate dynamic geological state interpretation maps.

[0050] The knowledge analysis submodule constructs a graph neural network constrained by a geological knowledge graph based on a dynamic geological state interpretation map; it applies Shapleyga and interpretation methods to quantify decision-making basis; and it calculates confidence intervals using Monte Carlo Dropout technology to generate quantitative risk assessment data.

[0051] Preferably, the twin decision support module includes a coupled modeling submodule, a data assimilation submodule, and a risk extrapolation submodule;

[0052] The coupled modeling submodule establishes a set of rock mass deformation-seepage control equations based on a three-dimensional geological property mesh model; initializes stress / seepage field boundary conditions; and generates a multiphysics coupled simulation base.

[0053] The data assimilation submodule, based on the multi-physics field coupled simulation base and the pre-processed signal dataset, uses an ensemble Kalman filter algorithm to dynamically invert permeability and modulus parameters, and generates a calibrated multi-field coupled model.

[0054] The risk simulation submodule, based on a calibrated multi-field coupling model and quantitative risk assessment data, injects engineering disturbance parameters; it predicts the evolution path of water inrush / rockburst through probabilistic finite element simulation and generates a dynamic multi-field coupling digital twin.

[0055] This invention provides an intelligent geological structure detection system and method for deep, special spaces. It offers the following advantages:

[0056] This invention utilizes synchronous acquisition via a distributed multiphysics sensing network and quantum-inspired adaptive filtering technology to effectively enhance weak feature signals in noisy environments, significantly improving the signal-to-noise ratio and the accuracy of target geological structure identification. Combined with a neural network inversion model constrained by physical information, the geophysical control equations are embedded as regularization terms in the loss function, overcoming the multi-solution dilemma of traditional inversion and outputting a high-confidence three-dimensional spatial distribution of physical property parameters. Based on an edge-cloud collaborative interpretation mechanism using a federated learning architecture, a lightweight model enables local real-time event detection, and combined with global model optimization and a background geological framework, achieves millisecond-level dynamic risk assessment response. Integrating a knowledge graph-guided graph neural network and interpretable analysis algorithms provides attribution basis and confidence interval quantification for key geological conclusions, enhancing decision-making transparency. A geological-stress-seepage multi-field coupled digital twin is constructed, and time-varying parameters are dynamically calibrated through data assimilation technology to predict the evolution path of engineering disturbance scenarios, supporting advanced disaster prevention and control. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the present invention;

[0058] Figure 2 This is a schematic diagram of the present invention;

[0059] Figure 3 This is a schematic diagram of the present invention;

[0060] Figure 4 This is a schematic diagram of the present invention;

[0061] Figure 5 This is a schematic diagram of the present invention;

[0062] Figure 6 This is a schematic diagram of the present invention;

[0063] Figure 7 This is a schematic diagram of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Example:

[0066] like Figure 1-7 As shown, this embodiment of the invention provides an intelligent geological structure detection system and method for deep, special spaces, comprising the following steps:

[0067] S1: Deploy a distributed sensor network in a deep, special space to simultaneously collect raw detection data of multiple physical fields, including seismic waves, electromagnetic fields, sound waves, temperature, and strain; construct a noise channel model based on quantum computing principles to model the raw detection data; apply a quantum-inspired adaptive filtering algorithm to estimate noise characteristics in real time through variational optimization and dynamically adjust filtering parameters to process the raw detection data, thereby achieving strong background noise suppression and enhancement of weak feature signals related to the target geological structure, and generating a preprocessed signal dataset;

[0068] S2: Based on the preprocessed signal dataset, depth features of each physical field data are extracted using a convolutional neural network model; a feature fusion module based on an attention mechanism is used to dynamically weight and fuse the depth features to generate a joint feature representation; the joint feature representation is input into a neural network inversion model constrained by physical information, which uses the control equations describing the geophysical response as regularization terms embedded in the loss function to perform three-dimensional spatial geological physical parameter inversion and geological interface identification, generating a three-dimensional geological physical parameter model;

[0069] S3: Based on the background geological framework provided by the 3D geological physical parameter model and the real-time data stream in the preprocessed signal dataset; on the edge computing nodes deployed at the front end of the detection, a lightweight deep learning model is used to perform preliminary geological event detection and classification on the local real-time data; a federated learning architecture is constructed, where edge nodes only upload model updates to the central server, and the central server aggregates and updates to generate a global interpretation model and distributes it; edge nodes combine the local model, the global interpretation model, and the background geological framework to perform collaborative interpretation of the real-time data; based on the interpretation results and pre-set risk assessment rules, the geological hazard risk level is calculated in real time, generating real-time geological state interpretation data and risk level distribution data;

[0070] S4: Based on a three-dimensional geological physical parameter model, real-time geological state interpretation data and risk level distribution data, as well as a pre-set regional geological prior knowledge base; construct a knowledge-guided graph neural network model, and use geological structural framework constraint feature propagation and interpretation; apply interpretability analysis algorithms to analyze the decision basis of key models in S2 and S3; use uncertainty quantification technology to estimate the confidence level of key geological interface locations and risk level assessment results, and generate interpretable geological structure probability models and quantitative risk assessment data;

[0071] S5: Integrate a three-dimensional geological physical property parameter model, real-time geological state interpretation data and risk level distribution data, interpretable geological structure probability model and quantitative risk assessment data, and real-time multi-physics field data stream from the preprocessed signal dataset; establish a geological-stress-seepage multi-field coupled numerical simulation model; apply a data assimilation algorithm to dynamically invert and update the key time-varying parameters of the simulation model using the real-time multi-source data; construct a digital twin that integrates geological structure, real-time sensing state, and updated multi-field information; based on the digital twin, perform multi-physics field short-term evolution trend prediction and engineering disturbance scenario simulation and risk assessment, and generate a dynamic deep space geological-engineering multi-field coupled digital twin and prediction data.

[0072] S101: Deploy a distributed sensor network in a special deep space to synchronously collect raw seismic wave, electromagnetic field, sound wave, temperature and strain data to generate raw multiphysics time series datasets;

[0073] Seismic detector arrays are deployed at predetermined intervals on the sidewalls of the main roadway in deep mines, while distributed fiber optic sensors are installed on the roof and electromagnetic sensors are arranged in the boreholes. Multi-source data is synchronously acquired through an industrial ring network, with the seismic wave sampling rate set to 2000Hz, the electromagnetic field frequency band covering 100Hz-10kHz, and the strain data spatial resolution set to 5cm. A raw dataset containing timestamps, equipment identifiers, and physical quantity types is generated.

[0074] S102: Based on the original multiphysics time series dataset, using the principle of quantum computing, the signal is regarded as a "quantum state" and the noise is simulated as a "quantum noise channel", a hybrid signal-noise mathematical model is constructed, and a quantum noise channel model is generated;

[0075] Extract raw seismic wave data segments and normalize the amplitude into quantum state representations. ,in Corresponding base signal, Corresponding noise components; based on the measured electromagnetic interference spectrum characteristics, a decoherent noise channel model is constructed. Among them, the Kraus operator Determined by the noise power distribution; a quantum noise mathematical model containing a frequency band weight matrix is ​​generated.

[0076] S103: Based on the quantum noise channel model, the variational optimization algorithm is applied to iteratively estimate the type, intensity, and correlation parameters of noise in real time, and generate a dynamic noise characteristic parameter set;

[0077] Initialize a parameter matrix containing the standard deviation of Gaussian noise, impulse noise density, and correlation coefficient; define a variational loss function containing the fitting term and regularization term for the observed data; iteratively update the noise parameters using gradient descent, and terminate the optimization when the rate of change of the loss function is less than a predetermined threshold; output a dynamic noise characteristic parameter set.

[0078] S104: Based on the original multiphysics time series dataset and dynamic noise characteristic parameter set, a quantum-inspired adaptive filtering algorithm is adopted to dynamically adjust the filter parameters, suppress background noise and enhance the weak geological features of the target, and generate a feature enhancement signal.

[0079] A quantum-inspired filter with a center frequency matching the resonant frequency of the target geological body is constructed, and its phase compensation term is dynamically adjusted according to the noise parameters. A time-domain convolution operation is performed on the original data, in which the filter coefficients are updated in real time. After processing, the amplitude of the target signal is increased to more than 15 times the original value, and the noise standard deviation is reduced to less than 25% of the original value.

[0080] S105: Based on feature enhancement signals, timestamp synchronization and spatial interpolation algorithms are used to unify the spatiotemporal reference of different sensors, eliminate acquisition delay and spatial offset, and generate a preprocessed signal dataset.

[0081] Perform timestamp correction based on sensor position coordinates and clock offset data: The attenuation coefficient k is set according to the network delay; the non-uniformly distributed data is resampled to a standard grid using a spatial interpolation algorithm to generate a preprocessed signal dataset with a unified spatiotemporal reference.

[0082] S201: Based on the preprocessed signal dataset, an improved convolutional neural network is used to extract high-dimensional abstract features from seismic, electromagnetic, and acoustic data respectively, generating a single-modal deep feature set;

[0083] The preprocessed seismic wave data (time window 2 seconds, sampling points 4000) is input into a depth residual convolutional network. The first layer has a kernel size of 7×7 and a stride of 2 to extract low-frequency contour features. The second layer uses 3×3 hollow convolution with an expansion rate of 2 to capture large-scale geological interface reflection features. The third layer reduces the dimensionality through 1×1 convolution. In the example, an anomalous reflection wave with an amplitude of 0.35 and a frequency of 80Hz was identified at a burial depth of 800 meters, and a 128-dimensional feature vector was output.

[0084] S202: Based on a single-modal deep feature set, a gated attention mechanism is used to dynamically calculate the spatial weights of different modal features, perform cross-modal weighted fusion, and generate a multimodal joint feature tensor.

[0085] earthquake feature vectors Electromagnetic characteristics Sound wave characteristics Calculate attention score based on spatial alignment. ,in For global average pooling features, the weight matrix W is initialized as a random orthogonal matrix. In the example, at a certain coordinate point, the seismic feature score is 0.8, the electromagnetic feature score is 0.5, and the acoustic feature score is 0.6. The normalized weights... fusion features Generate a 256-dimensional joint feature tensor.

[0086] S203: Based on the multimodal joint feature tensor, a physical information constrained neural network is constructed. The elastic wave equation / Maxwell equation is embedded as a regularization term in the loss function, and the constrained inversion process conforms to physical laws, generating a physical regularized inversion model.

[0087] Construct an inversion network to output the velocity v(x,y,z) Density ρ(x,y,z) Resistivity R(x,y,z) loss function For observation data, This is the network prediction value. (Discrete form of elastic wave equation), with the physical constraint coefficient λ set to 0.7. In the example, in the fault region... The term reduces the velocity gradient outlier from 0.25 to 0.08.

[0088] S204: Based on the physical regularization inversion model, the gradient descent optimization algorithm is used to iteratively update physical property parameters such as velocity, density, and resistivity until the loss function converges, generating a three-dimensional geological physical property parameter grid model.

[0089] Initialize the physical property parameter matrix Calculate the gradient of the loss function ,renew Learning rate The gradient norm is calculated at each step of the iteration. When 5 consecutive steps The process terminates at a certain point; in this example, after the 120th iteration, v is updated to 3850 m / s. Updated to 2.8 g / cm³, outputting a physical property parameter mesh with a resolution of 0.5 meters.

[0090] S301: Deploy lightweight convolutional networks with knowledge distillation and compression on edge devices such as mining robots and sensor nodes to support real-time data processing and generate edge intelligent interpretation models;

[0091] A Jetson TX2 computing unit was built into a mining inspection robot, loaded with a MobileNetV3 network compressed by knowledge distillation, with the original teacher model being ResNet50. The KL divergence loss was minimized. With a temperature coefficient T=5, the parameter size was compressed from 23.5MB to 4.8MB, and the inference latency on a 1.2GHz CPU in the instance was reduced to 15ms, meeting the 200ms real-time response requirement.

[0092] S302: Based on the preprocessed signal dataset and the edge intelligent interpretation model, microseismic event localization and electromagnetic anomaly classification are performed at the edge, generating a local geological event label set;

[0093] Edge nodes acquire preprocessed microseismic signals (time window 50ms), inputting them into a lightweight model for event detection: convolutional layers extract time-frequency features, and fully connected layers output event type probabilities. When P(rock crack) Record the labels {coordinates (x, y, z)} when the seismic velocity is greater than 0.9 and the source location error is less than 3m. Type, Energy}, In the example, a microseismic signal (dominant frequency 120Hz, duration 0.8s) at a certain tunnel face is labeled as coordinates (235, 178, -812). Type A rock fracture.

[0094] S303: Based on a local geological event tag set, a federated learning framework is adopted, with edge nodes uploading encrypted model gradients to the central server; the server generates a global model through a secure aggregation algorithm, and generates a federated optimized global interpretation model;

[0095] Each node calculates the local model gradient. Add Laplace noise After encryption, the data is uploaded, and the central server performs secure aggregation. When there are ≥10 participating nodes and the gradient variance is <0.05, the global model is updated. In the example, after the 20th aggregation, the microseismic identification accuracy increased from 81% to 93%, and a new model parameter file was generated.

[0096] S304: Based on a federated optimization global interpretation model, a three-dimensional geological property parameter model, and a local geological event tag set, data is fused at the edge to identify changes in rock mass stability and geological anomalies, and to generate a dynamic geological state interpretation map.

[0097] The edge device invokes local geological event tags {fault activation events} and overlays the velocity gradient field from the 3D physical property model. (Threshold > 200 m / s / m is considered an outlier zone), using decision tree rules: if If the frequency of events is greater than 5 times per minute, it is marked as "rock mass instability". In the example, the coordinates are (240, 180, -815). A velocity gradient of 285 m / s / m with an event frequency of 8 times / minute is generated, triggering a red warning zone.

[0098] S305: Based on the dynamic geological state interpretation map, a random forest risk assessment model is adopted, combined with a pre-set rule base, to output spatial risk levels and generate real-time geological state interpretation and risk distribution data.

[0099] Construct a random forest model: Input the outlier area, mean event energy, and property gradient values ​​from the interpretation graph in the input layer; the decision tree node splitting rule is Gini exponent < 0.2; output the risk probability. Set threshold The area in the example is considered high-risk. The physical property gradient abrupt change value reached 310 m / s / m, which is marked as a high-risk level.

[0100] S401: Integrating regional borehole data, structural maps, and lithological sequences, and using graph database technology, a geological prior knowledge graph containing stratigraphic contact relationships, fault attributes, and physical property ranges is constructed to generate a structured geological knowledge base;

[0101] Core logging data from 20 boreholes in the target mining area were collected, and stratigraphic codes, burial depths, and lithological codes were extracted. A "Stratigraphic" node was created in the Neo4j graph database, with attributes including average density of 2.65–2.85 g / cm³ and wave velocity range of 3500–4500 m / s. A "Fault" node was established based on the geological structure map, with attributes set to strike 120°, dip 65°, and displacement 8.5 m. Adjacent stratigraphic nodes were connected via "Contact Relationship" edges, and the "Overlying" relationship type was defined. In the example, the Permian P2h node was connected to the Carboniferous C2b node via an "Angular Unconformity" edge, generating a structured knowledge base containing 357 nodes and 428 edges.

[0102] S402: Based on a three-dimensional geological physical property parameter model, real-time geological state interpretation and risk distribution data, and a structured geological knowledge base, a graph neural network is constructed to generate knowledge-guided geological structure interpretation results by constraining feature propagation through knowledge graph node relationships.

[0103] Velocity field of three-dimensional physical property model Discretize the graph into a 50×50×30 grid, with each grid point serving as a graph node, and initialize the node features. Retrieve the spatial location of the fault from the knowledge base and add "construction constraint" edges to the corresponding grid; Use a two-layer graph convolutional network: ,in Given a constrained adjacency matrix, These are trainable weights; in the example, after propagation, the feature vector of a node in a fault zone is updated from [3850,0,0.2] to [3820,215,0.7], and the output is a spatial distribution map of the marked fault.

[0104] S403: Based on the knowledge-guided geological structure interpretation results, the contribution of input features to fault identification decisions is quantified using the Shapley sum interpretation method or gradient salience map, and a geological interpretation attribution report is generated.

[0105] Select coordinates (120, 80, -500) from the interpretation results. Fault identification points, calculate Shapley values: construct input feature set Marginal contribution is calculated using perturbation characteristics. ,in The model outputs probabilities for subset S, in the example. Generate a feature contribution ranking report.

[0106] S404: Based on the knowledge-guided geological structure interpretation results, Monte Carlo Dropout technology is used to perform multiple sampling predictions on the locations of key geological interfaces, calculate confidence intervals, and generate geological structure confidence interval data.

[0107] For key geological interface nodes (120, 80, -500) Perform 50 Monte Carlo Dropout predictions: randomly discard 20% of the neural network nodes during each forward propagation, and record the depth prediction values. ; Calculate the mean Standard deviation Determine the 95% confidence interval. Output depth value range data.

[0108] S405: Integrates geological interpretation attribution reports, geological structure confidence interval data, real-time geological status interpretation and risk distribution data, generates a risk assessment with confidence level, and generates an interpretable geological structure probability model and quantitative risk assessment data.

[0109] Construct a Bayesian network: the parent node contains "probability of fault existence" (prior value 0.85) and "risk level" (high / medium / low); define a conditional probability table P (water inrush | fault, risk). When a fault exists and is of high risk, P=0.92; input attribution report Given a confidence interval half-width of 1.96σ = 0.78m and a real-time risk value of 0.7, calculate the posterior probability. A threshold of 0.5 is set to trigger an early warning and generate quantitative risk assessment data.

[0110] S501: Based on the three-dimensional geological physical property parameter model, establish the rock mass deformation-seepage control equation, initialize the stress field and seepage field parameters, and generate a multi-physics field coupled simulation model.

[0111] Density field in a three-dimensional physical property parameter model With velocity field Calculate the initial stress field (Elastic tensor C is converted from v, Poisson's ratio ν = 0.25), seepage field initialization: permeability k = 10mD, porosity pressure gradient In coordinates (200, 150, -800) Establish the governing equations: equilibrium equations seepage continuity equation ,in This generates a multi-field coupled model containing 12,000 grid cells.

[0112] S502: Based on a multi-physics field coupled simulation model, preprocessed signal dataset, real-time geological state interpretation and risk distribution data, a multi-physics field coupled model is generated by dynamically inverting and updating permeability and rock mass modulus parameters using ensemble Kalman filtering.

[0113] Construct a set of 100 members: apply a ±15% perturbation to the permeability k and a ±10% perturbation to the modulus E, and take real-time strain data from the preprocessed dataset. Predicting strain Calculate the Kalman gain using the coupled model. The observation noise covariance R = diag(0.0001^2) Update parameters In the example, in a low-permeability zone, k was updated from 8.5 mD to 11.2 mD, and E was updated from 25 GPa to 22.8 GPa.

[0114] S503: Based on a calibrated multi-field coupling model, an interpretable geological structure probability model, and quantitative risk assessment data, a three-dimensional mesh model containing parameter uncertainties is constructed to generate a probabilistic digital twin base.

[0115] Model the penetration rate k as a Gaussian random field: ,in Taken from the calibration model, Calculated from confidence interval data (corresponding to half-width 0.8mD) ), in the fault zone (235, 178, -812) Set the spatial correlation length to 50m, generate a probabilistic parameter field, and integrate risk data. The area is marked as a red warning unit, and the output is a twin base model containing 785 risk units.

[0116] S504: Based on a probabilistic digital twin base, using finite element numerical simulation, it calculates the stress field redistribution and seepage path expansion trend in the next 24 hours, and generates multiphysics short-term prediction data.

[0117] Set boundary conditions: Apply mining stress increment to the area 20m in front of the tunnel face. Solve the finite element equations: stiffness matrix Percolation matrix Time step Minutes, iterative calculations over 24 hours (144 steps), recording the maximum principal stress. change: The seepage velocity increased from 0.15 m / d to 0.38 m / d, and the predicted data file was output.

[0118] S505: Based on a probabilistic digital twin base, it injects engineering disturbance parameters, simulates the evolution of surrounding rock damage and water inrush risk, outputs risk warning maps, and generates dynamic geological-engineering multi-field coupled digital twins and prediction data.

[0119] Injection excavation parameters: cross-section advance speed 2m / d, support delay 6 hours, in the damage model Set intensity threshold Shape parameter m=3, simulated water inrush criterion: when An alert is triggered when D > 0.7, in the example coordinates (240, 180, -815). At t=18h, p=0.28MPa and D=0.75, it is marked as a high-risk area for water inrush.

[0120] A smart geological structure detection system for deep and special spaces includes the following modules: a multi-source sensing and processing module, which uses a quantum-inspired adaptive filtering algorithm to suppress noise in raw seismic wave, electromagnetic field, and acoustic wave data based on a deployed distributed sensor network; unifies the multi-source data benchmark through a spatiotemporal alignment algorithm; and outputs an enhanced standardized data stream to generate a preprocessed signal dataset.

[0121] The intelligent interpretation and assessment module, based on a preprocessed signal dataset, uses a physical information-constrained neural network to invert 3D geological parameters; it achieves edge-cloud collaborative interpretation through a federated learning framework; and it combines a knowledge-guided graph neural network to output interpretable risk assessments, generating 3D geological property parameter models and quantitative risk assessment data.

[0122] The twin decision support module constructs a rock mass deformation-seepage coupling model based on a three-dimensional geological physical parameter model and quantitative risk assessment data; it uses an ensemble Kalman filter algorithm to dynamically assimilate real-time data; and it generates engineering disturbance response predictions through a probabilistic twin engine, thus generating a dynamic multi-field coupled digital twin.

[0123] The multi-source sensing processing module includes a heterogeneous acquisition submodule, a quantum noise reduction submodule, and a spatiotemporal alignment submodule;

[0124] The heterogeneous acquisition submodule deploys arrays of seismic, electromagnetic, and acoustic sensors to simultaneously acquire raw physical field data from deep space and generate raw multiphysics datasets.

[0125] A three-component seismic detector array was deployed on the smooth surface of the mine roadway with a horizontal spacing of 10 meters. Distributed fiber optic strain sensors (spatial resolution of 5 cm) were installed on the roof, and broadband electromagnetic sensors (100Hz-10kHz) were embedded in the borehole. The synchronization trigger signal was provided by a GPS disciplined clock with a time synchronization accuracy of ±1μs. In the example, the seismic waveform amplitude was 0.25g, the strain value was 12με, and the electromagnetic intensity was 0.3V / m during a certain acquisition cycle, generating a multiphysics data matrix with a unified timestamp.

[0126] The quantum noise reduction submodule, based on the original multiphysics dataset, constructs a hybrid model using a quantum noise channel modeling algorithm; dynamically estimates noise parameters using a variational optimization algorithm; and suppresses noise and generates a feature-enhanced signal set through a quantum-inspired adaptive filter.

[0127] Extract raw seismic data segments (time window 0.5-1.0 seconds), normalize the amplitude to construct quantum states |ψ>, where the base amplitude |0> corresponds to the theoretical waveform and the noise component |1> corresponds to the measured deviation; construct the Kraus operator based on the main peak of the electromagnetic background spectrum 3000Hz±200Hz. , Frequency band power ratio (at 3000Hz) Initialize the noise parameter vector. (Initial value of Gaussian noise standard deviation σ: 0.15), define the loss function. Gradient descent update (η=0.01), after iteration, σ converges to 0.08; a bandpass filter is constructed. Center frequency 120Hz, bandwidth 40Hz, for raw data Perform convolution In the example The amplitude of the target signal increased from 0.02 to 0.31.

[0128] The spatiotemporal alignment submodule, based on the feature-enhanced signal set, uses a timestamp synchronization algorithm to eliminate acquisition delay; it unifies the spatial reference through Kriging spatial interpolation and generates a preprocessed signal dataset.

[0129] Detect sensor clock deviation, maximum time offset 3ms, and perform compensation: Reference node ; Obtain the sensor's three-dimensional coordinates and identify missing grid points Perform Kriging interpolation: In the example, the data from the 8-meter-spaced detectors are interpolated to a 1-meter grid to generate a data cube with a unified spatiotemporal reference.

[0130] The intelligent interpretation and evaluation module includes a property inversion submodule, a federated interpretation submodule, and a knowledge analysis submodule;

[0131] The physical property inversion submodule, based on a preprocessed signal dataset, uses a deep residual convolutional network to extract single-modal features; it achieves cross-modal fusion through a gated attention mechanism; it constructs a physically constrained inversion model to output three-dimensional physical property parameters and generates a three-dimensional geological physical property mesh model.

[0132] Preprocessed seismic signals (2-second time window, 4000 sampling points) are input into a three-layer convolutional network: the first layer uses 7×7 convolutional kernels to extract low-frequency features, the second layer uses 3×3 dilated convolutions (dilation rate 2) to capture interface reflections, and the third layer uses 1×1 convolutions to reduce dimensionality. In this example, an anomalous wave with an amplitude of 0.35 and a frequency of 80Hz is identified at a burial depth of 800 meters, and a 128-dimensional feature vector is output. The seismic feature vector is then processed... Electromagnetic characteristics Sound wave characteristics Align by coordinates and calculate attention score. ( (for global features), a certain point Normalized weights fusion features ; Construct an inversion network to output speed loss function ( gradient descent update (Learning rate 0.01), after 120 iterations The speed converged from 3500 m / s to 3850 m / s, generating a 0.5-meter resolution property grid.

[0133] The federated interpretation submodule is based on a 3D geological property mesh model and deploys a lightweight convolutional network at the edge. It uses a secure aggregation federated learning framework to update the global model and combines real-time data streams to identify geological anomalies and generate dynamic geological state interpretation maps.

[0134] A MobileNetV3 network (4.8MB of parameters) is deployed in the embedded computing unit of a mine inspection robot. Real-time strain data (sampling rate 100Hz) is input, and the output rock fracture probability is calculated. And positioning error Generate label in meters {coordinates (235, 178, -812)} Type A rock fracture}; Local calculation model gradient Add Laplace noise Encrypted upload, centralized server aggregation (Participating nodes) And gradient variance After the update, the accuracy of global model microseismic identification improved from 81% to 93%; the velocity gradient of the physical property model was integrated at the edge. (Threshold 200m / s / m), if And the frequency of events The frequency of incidents per minute is labeled "rock mass instability". In the example, the coordinates are (240, 180, -815). Place The overlay frequency is 8 times per minute, and the warning area map is output.

[0135] The knowledge analysis submodule constructs a graph neural network constrained by a geological knowledge graph based on a dynamic geological state interpretation map; it applies Shapleyga and interpretation methods to quantify decision-making basis; and it calculates confidence intervals using Monte Carlo Dropout technology to generate quantitative risk assessment data.

[0136] Constructing a Graph Neural Network: Features of Material Grid Nodes Add constraint edges to knowledge base fault nodes, and perform graph convolution propagation. ( To constrain the adjacency matrix, the features of a fault node are updated from [3850, 0, 0.2] to [3820, 215, 0.7]; the coordinates (120, 80, -500) are selected. Fault points, calculate Shapley values In the example Perform 50 Monte Carlo Dropout predictions, depth value Standard deviation meters, 95% confidence interval Input Bayesian network computation (Threshold 0.5) Generate quantitative risk assessment data.

[0137] The twin decision support module includes a coupled modeling submodule, a data assimilation submodule, and a risk extrapolation submodule;

[0138] The coupled modeling submodule establishes a set of rock mass deformation-seepage control equations based on a three-dimensional geological property mesh model; initializes stress / seepage field boundary conditions; and generates a multiphysics coupled simulation base.

[0139] The velocity field at a depth of 800 meters in the three-dimensional physical property mesh model is taken. m / s, density field g / cm³, calculate the elastic tensor using Hooke's law. Young's modulus Poisson's ratio ,have to (GPa), establish equilibrium equations (physical strength Taking the directional component of gravitational acceleration (9.8 m / s²), the seepage field is initialized as follows: permeability. mD, porosity pressure gradient MPa / m, continuity equation ( Dynamic viscosity Pa·s), applying mining stress increment in the 20-meter area ahead of the tunnel face. MPa, generating a coupled simulation base containing 15,000 hexahedral elements.

[0140] The data assimilation submodule, based on the multi-physics field coupled simulation base and the pre-processed signal dataset, uses an ensemble Kalman filter algorithm to dynamically invert permeability and modulus parameters, and generates a calibrated multi-field coupled model.

[0141] Construct 100 set members: for penetration rate Apply random perturbation ( mD), modulus Apply Disturbance ( GPa), take the preprocessed dataset Strain observations at different times Each member calculates and predicts strain. Calculate the covariance matrix Observation noise covariance (Based on sensor accuracy), Kalman gain Update parameters In an example, a low-permeability area Updated from 9.8mD to 12.5mD. Updated from 40.2 GPa to 36.8 GPa.

[0142] The risk simulation submodule, based on a calibrated multi-field coupling model and quantitative risk assessment data, injects engineering disturbance parameters; it predicts the evolution path of water inrush / rockburst through probabilistic finite element simulation and generates a dynamic multi-field coupling digital twin.

[0143] Injection engineering parameters: excavation cross-section size 5m×4m, daily advance 2.5m, support delay 4 hours, damage model set. (Intensity threshold) MPa is based on the uniaxial compressive strength test of rock and shape parameters. Criterion for sudden water inrush: When the pore water pressure ( The height of the crack. MPa is the osmotic pressure resistance threshold), and in the finite element model, risk elements ( Apply the damage evolution equation, time step Minutes, simulating 24 hours (288 steps), recording coordinates ( Place hour MPa ; hour MPa Triggering the warning threshold Output a path diagram of the evolution of water inrush risk.

[0144] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent detection of geological structures in deep, special spaces, characterized in that, Includes the following steps: S1: Deploy a distributed sensor network in deep, special spaces to simultaneously collect raw detection data of multiple physical fields, including seismic waves, electromagnetic fields, sound waves, temperature, and strain; A noise channel model is constructed based on quantum computing principles to model the original detection data. A quantum-inspired adaptive filtering algorithm is applied to process the original detection data by real-time estimation of noise characteristics through variational optimization and dynamic adjustment of filtering parameters. This process achieves strong background noise suppression and enhancement of weak feature signals related to the target geological structure, generating a preprocessed signal dataset. S2: Based on the preprocessed signal dataset, use a convolutional neural network model to extract the depth features of each physical field data; An attention-based feature fusion module is used to dynamically weight and fuse the deep features to generate a joint feature representation. The joint feature representation is then input into a physical information-constrained neural network inversion model. This model uses the control equations describing the geophysical response as regularization terms embedded in the loss function to perform three-dimensional spatial geological property parameter inversion and geological interface identification, thereby generating a three-dimensional geological property parameter model. S3: The background geological framework provided by the three-dimensional geological physical property parameter model, and the real-time data stream in the preprocessed signal dataset; On edge computing nodes deployed at the front end of the detection system, a lightweight deep learning model is used to perform preliminary geological event detection and classification on local real-time data. A federated learning architecture is constructed, where edge nodes only upload model updates to the central server. The central server aggregates and updates the data to generate a global interpretation model and distributes it. Edge nodes combine the local model, the global interpretation model, and the background geological framework to perform collaborative interpretation of real-time data. Based on the interpretation results and pre-set risk assessment rules, the geological hazard risk level is calculated in real time, generating real-time geological status interpretation data and risk level distribution data. S4: Based on a three-dimensional geological physical parameter model, real-time geological state interpretation data and risk level distribution data, as well as a pre-set regional geological prior knowledge base; construct a knowledge-guided graph neural network model, and use geological structural framework constraint feature propagation and interpretation; apply interpretability analysis algorithms to analyze the decision basis of key models in S2 and S3; use uncertainty quantification technology to estimate the confidence level of key geological interface locations and risk level assessment results, and generate interpretable geological structure probability models and quantitative risk assessment data; S5: Integrates three-dimensional geological physical parameter models, real-time geological state interpretation data and risk level distribution data, interpretable geological structure probability models and quantitative risk assessment data, as well as real-time multiphysics data streams in preprocessed signal datasets; A multi-field coupled numerical simulation model of geology, stress, and seepage was established; a data assimilation algorithm was applied to dynamically invert and update the key time-varying parameters of the simulation model using the real-time multi-source data. Construct a digital twin that integrates geological structure, real-time sensing status, and updated multi-field information; based on the digital twin, perform short-term evolution trend prediction of multi-physics fields and simulation and risk assessment of engineering disturbance scenarios, and generate a dynamic deep space geology-engineering multi-field coupled digital twin and prediction data.

2. The intelligent geological structure detection method for deep special spaces according to claim 1, characterized in that: S1 includes the following steps: S101: Deploy a distributed sensor network in a special deep space to synchronously collect raw seismic wave, electromagnetic field, sound wave, temperature and strain data to generate raw multiphysics time series datasets; S102: Based on the original multiphysics time series dataset, using the principle of quantum computing, the signal is regarded as a "quantum state" and the noise is simulated as a "quantum noise channel", a hybrid signal-noise mathematical model is constructed, and a quantum noise channel model is generated; S103: Based on the quantum noise channel model, the variational optimization algorithm is applied to iteratively estimate the type, intensity, and correlation parameters of noise in real time, and generate a dynamic noise characteristic parameter set; S104: Based on the original multiphysics time series dataset and dynamic noise characteristic parameter set, a quantum-inspired adaptive filtering algorithm is adopted to dynamically adjust the filter parameters, suppress background noise and enhance the weak geological features of the target, and generate a feature enhancement signal. S105: Based on feature enhancement signals, timestamp synchronization and spatial interpolation algorithms are used to unify the spatiotemporal reference of different sensors, eliminate acquisition delay and spatial offset, and generate a preprocessed signal dataset.

3. The intelligent geological structure detection method for deep special spaces according to claim 1, characterized in that: S2 includes the following steps: S201: Based on the preprocessed signal dataset, an improved convolutional neural network is used to extract high-dimensional abstract features from seismic, electromagnetic, and acoustic data respectively, generating a single-modal deep feature set; S202: Based on a single-modal deep feature set, a gated attention mechanism is used to dynamically calculate the spatial weights of different modal features, perform cross-modal weighted fusion, and generate a multimodal joint feature tensor. S203: Based on the multimodal joint feature tensor, a physical information constrained neural network is constructed. The elastic wave equation / Maxwell equation is embedded as a regularization term in the loss function, and the constrained inversion process conforms to physical laws, generating a physical regularized inversion model. S204: Based on the physical regularization inversion model, the gradient descent optimization algorithm is used to iteratively update physical property parameters such as velocity, density, and resistivity until the loss function converges, generating a three-dimensional geological physical property parameter grid model.

4. The intelligent geological structure detection method for deep special spaces according to claim 1, characterized in that: S3 includes the following steps: S301: Deploy lightweight convolutional networks with knowledge distillation and compression on edge devices such as mining robots and sensor nodes to support real-time data processing and generate edge intelligent interpretation models; S302: Based on the preprocessed signal dataset and the edge intelligent interpretation model, microseismic event localization and electromagnetic anomaly classification are performed at the edge, generating a local geological event label set; S303: Based on a local geological event tag set, a federated learning framework is adopted, with edge nodes uploading encrypted model gradients to the central server; the server generates a global model through a secure aggregation algorithm, and generates a federated optimized global interpretation model; S304: Based on a federated optimization global interpretation model, a three-dimensional geological property parameter model, and a local geological event tag set, data is fused at the edge to identify changes in rock mass stability and geological anomalies, and to generate a dynamic geological state interpretation map. S305: Based on the dynamic geological state interpretation map, a random forest risk assessment model is adopted, combined with a pre-set rule base, to output spatial risk levels and generate real-time geological state interpretation and risk distribution data.

5. The intelligent geological structure detection method for deep special spaces according to claim 1, characterized in that: S4 includes the following steps: S401: Integrating regional borehole data, structural maps, and lithological sequences, and using graph database technology, a geological prior knowledge graph containing stratigraphic contact relationships, fault attributes, and physical property ranges is constructed to generate a structured geological knowledge base; S402: Based on a three-dimensional geological physical property parameter model, real-time geological state interpretation and risk distribution data, and a structured geological knowledge base, a graph neural network is constructed to generate knowledge-guided geological structure interpretation results by constraining feature propagation through knowledge graph node relationships. S403: Based on the knowledge-guided geological structure interpretation results, the contribution of input features to fault identification decisions is quantified using the Shapley sum interpretation method or gradient salience map, and a geological interpretation attribution report is generated. S404: Based on the knowledge-guided geological structure interpretation results, Monte Carlo Dropout technology is used to perform multiple sampling predictions on the locations of key geological interfaces, calculate confidence intervals, and generate geological structure confidence interval data. S405: Integrates geological interpretation attribution reports, geological structure confidence interval data, real-time geological status interpretation and risk distribution data, generates a risk assessment with confidence level, and generates an interpretable geological structure probability model and quantitative risk assessment data.

6. The intelligent geological structure detection method for deep special spaces according to claim 1, characterized in that... based on: S5 includes the following steps: S501: Based on the three-dimensional geological physical property parameter model, establish the rock mass deformation-seepage control equation, initialize the stress field and seepage field parameters, and generate a multi-physics field coupled simulation model. S502: Based on a multi-physics field coupled simulation model, preprocessed signal dataset, real-time geological state interpretation and risk distribution data, a multi-physics field coupled model is generated by dynamically inverting and updating permeability and rock mass modulus parameters using ensemble Kalman filtering. S503: Based on a calibrated multi-field coupling model, an interpretable geological structure probability model, and quantitative risk assessment data, a three-dimensional mesh model containing parameter uncertainties is constructed to generate a probabilistic digital twin base. S504: Based on a probabilistic digital twin base, using finite element numerical simulation, it calculates the stress field redistribution and seepage path expansion trend in the next 24 hours, and generates multiphysics short-term prediction data. S505: Based on a probabilistic digital twin base, it injects engineering disturbance parameters, simulates the evolution of surrounding rock damage and water inrush risk, outputs risk warning maps, and generates dynamic geological-engineering multi-field coupled digital twins and prediction data.

7. An intelligent geological structure detection system for deep, special spaces, characterized in that, It includes the following modules: a multi-source sensing processing module, which uses a quantum-inspired adaptive filtering algorithm to suppress noise in raw seismic wave, electromagnetic field, and acoustic wave data based on a deployed distributed sensor network; it unifies the multi-source data benchmark through a spatiotemporal alignment algorithm; and it outputs an enhanced standardized data stream to generate a preprocessed signal dataset. The intelligent interpretation and evaluation module, based on a preprocessed signal dataset, uses a physical information-constrained neural network to invert three-dimensional geological parameters; Enables edge-cloud collaborative interpretation through a federated learning framework; By combining the knowledge-guided graph neural network output with interpretable risk assessment, a three-dimensional geological physical parameter model and quantitative risk assessment data are generated. The twin decision support module constructs a rock mass deformation-seepage coupling model based on a three-dimensional geological physical property parameter model and quantitative risk assessment data; and uses an ensemble Kalman filter algorithm to dynamically assimilate real-time data. By generating engineering disturbance response predictions through a probabilistic twin engine, a dynamic multi-field coupled digital twin is generated.

8. The intelligent geological structure detection system for deep special spaces according to claim 7, characterized in that: The multi-source sensing processing module includes a heterogeneous acquisition submodule, a quantum noise reduction submodule, and a spatiotemporal alignment submodule; The heterogeneous acquisition submodule deploys arrays of seismic, electromagnetic, and acoustic sensors to synchronously acquire raw physical field data of deep space and generate raw multiphysics datasets. The quantum noise reduction submodule, based on the original multiphysics dataset, constructs a hybrid model using a quantum noise channel modeling algorithm; dynamically estimates noise parameters using a variational optimization algorithm; and suppresses noise and generates a feature-enhanced signal set through a quantum-inspired adaptive filter. The spatiotemporal alignment submodule, based on the feature-enhanced signal set, uses a timestamp synchronization algorithm to eliminate acquisition delay; A preprocessed signal dataset is generated by unifying the spatial reference through Kriging space interpolation.

9. The intelligent geological structure detection system for deep special spaces according to claim 7, characterized in that: The intelligent interpretation and evaluation module includes a property inversion submodule, a federated interpretation submodule, and a knowledge analysis submodule; The physical property inversion submodule, based on a preprocessed signal dataset, uses a deep residual convolutional network to extract single-modal features; it achieves cross-modal fusion through a gated attention mechanism; it constructs a physically constrained inversion model to output three-dimensional physical property parameters and generates a three-dimensional geological physical property mesh model. The federated interpretation submodule is based on a 3D geological property mesh model and deploys a lightweight convolutional network at the edge. It uses a secure aggregation federated learning framework to update the global model and combines real-time data streams to identify geological anomalies and generate dynamic geological state interpretation maps. The knowledge analysis submodule constructs a graph neural network constrained by a geological knowledge graph based on a dynamic geological state interpretation map; it applies Shapleyga and interpretation methods to quantify decision-making basis; and it calculates confidence intervals using Monte Carlo Dropout technology to generate quantitative risk assessment data.

10. The intelligent geological structure detection system for deep special spaces according to claim 7, characterized in that: The twin decision support module includes a coupled modeling submodule, a data assimilation submodule, and a risk extrapolation submodule; The coupled modeling submodule establishes a set of rock mass deformation-seepage control equations based on a three-dimensional geological property grid model. Initialize stress / seepage field boundary conditions to generate a multiphysics coupled simulation base; The data assimilation submodule, based on the multi-physics field coupled simulation base and the pre-processed signal dataset, uses an ensemble Kalman filter algorithm to dynamically invert permeability and modulus parameters, and generates a calibrated multi-field coupled model. The risk simulation submodule, based on a calibrated multi-field coupling model and quantitative risk assessment data, injects engineering disturbance parameters; The evolution path of water inrush / rockburst is predicted by probabilistic finite element simulation, and a dynamic multi-field coupled digital twin is generated.

Images