A survey method for identifying a coal mine goaf

Abstract

This application relates to the field of geophysical exploration technology and discloses an exploration method for identifying coal mine goaf areas. The method includes acquiring multi-source geophysical data of the exploration area; spatially aligning and gridding the data; and using a deep autoencoder network for data-driven feature fusion to construct a unified feature space. Then, a structure-coupled collaborative inversion is performed, and by constructing and minimizing a joint objective function containing cross-gradient structure coupling terms, a multi-physical attribute model that maintains spatial consistency is obtained. Furthermore, a three-dimensional convolutional neural network is used to perform voxel-level classification of the inversion results, automatically interpreting and generating a three-dimensional geological classification model indicating the location and attributes of the goaf. This invention, through data-driven fusion, structure-coupled inversion, and intelligent interpretation, effectively reduces the ambiguity of inversion, replaces the subjective manual interpretation process, and significantly improves the accuracy, efficiency, and objectivity of goaf detection.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, specifically to an exploration method for identifying coal mine goaf areas. Background Technology

[0002] The goaf formed after coal mining and the disturbance and damage to the overlying strata pose significant safety hazards to mine production and the surface environment. Accurately determining the spatial distribution, morphology, scale, and internal filling status of old goaf areas is crucial for preventing disasters such as water, fire, and gas, as well as for the rational planning of surface engineering construction. Geophysical exploration is an effective technical means for long-distance, non-destructive investigation of underground geological structures and the spatial distribution of goaf areas.

[0003] In practical applications, single geophysical methods, such as transient electromagnetic methods, direct current methods, or seismic exploration methods, are often used for detection. However, due to the complexity of underground geological conditions and the inherent ambiguity of geophysical inversion, the same geophysical anomaly may be caused by different geological bodies. For example, resistivity anomalies may correspond to dry goaf areas or be caused by other high-resistivity geological bodies. Therefore, it is difficult to accurately define the boundaries and internal conditions of goaf areas based solely on anomalies of a single physical property.

[0004] To overcome the limitations of single methods, technicians have attempted to integrate multiple exploration methods in order to make judgments based on comprehensive information from multiple physical attributes. However, at the data processing and interpretation level, existing fusion methods mostly remain at the level of simple superposition and comparison of the inversion results from different methods. This approach fails to achieve deep data fusion; no inherent constraints are established between data from different sources and with different physical properties, resulting in a highly subjective comprehensive interpretation process and making it difficult to effectively utilize the inherent correlations between multi-source data.

[0005] Furthermore, the current interpretation of exploration results largely relies on manual delineation and experience-based judgment by professionals. This method is not only inefficient, but the accuracy and objectivity of the interpretation results are also greatly affected by subjective factors, making it difficult to meet the needs for a refined and quantitative description of goaf areas.

[0006] Therefore, developing a method that can perform deep fusion and automated joint interpretation of multi-source geophysical data to improve the accuracy and reliability of goaf detection is of great practical significance and application value. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for identifying coal mine goaf areas, which solves the problems of low detection accuracy, high subjectivity, and low efficiency in existing goaf area exploration methods that rely on a single geophysical exploration method and manual interpretation.

[0008] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides an exploration method for identifying coal mine goaf areas. This method improves the accuracy and automation level of goaf area identification by combining multi-source data fusion, structural coupling collaborative inversion and intelligent algorithm interpretation.

[0009] The technical solution provided by the first aspect of this invention specifically includes the following steps: S1: Obtain at least two types of geophysical exploration data from the exploration area; S2: Perform data preprocessing on the geophysical exploration data; S3: Establish a three-dimensional grid covering the exploration area, and map the preprocessed physical attribute data to the nodes of the three-dimensional grid through spatial interpolation to form a high-dimensional physical attribute vector; S4: Based on the preset cross-modal data fusion model, the high-dimensional physical attribute vector is converted into a fusion feature vector in a unified feature space; S5: Through collaborative inversion and intelligent interpretation, a three-dimensional geological classification model representing the spatial location and attributes of the goaf is generated; S6: Perform three-dimensional visualization on the three-dimensional geological classification model to generate a three-dimensional visualization model of the goaf.

[0010] As a preferred embodiment, the cross-modal data fusion model is a deep autoencoder network. This network comprises an encoder and a decoder. The high-dimensional physical property vector... Input encoder Low-dimensional fused feature vectors are obtained through nonlinear transformation compression. decoder Then try to fuse the feature vectors Reconstructed into the original physical property vector The network minimizes reconstruction errors (e.g., mean squared error). Once trained, the encoder will have the ability to map multiple physical attribute data to a unified feature space.

[0011] As a preferred implementation, the cooperative inversion is achieved by constructing and minimizing a joint objective function. To achieve this. For the two physical property models to be inverted. and The function can be expressed as: ; in, For data fitting terms, For model regularization terms, For structural coupling terms, , , This is the regularization parameter. This is the structure coupling term. This can be achieved through a cross-gradient function, which is defined as the integral of the magnitude of the cross product of the gradient vector in the model space. By minimizing this joint objective function, the different physical property models obtained through inversion can be made to maintain consistency in spatial structure.

[0012] In a preferred embodiment, the intelligent interpretation is achieved through a pre-trained intelligent interpretation model, which performs voxel-level classification on the 3D multi-physical attribute model obtained by collaborative inversion. This intelligent interpretation model can be a 3D convolutional neural network model. Its training process includes: constructing a training sample set consisting of 3D data blocks and their geological labels using known geological information such as borehole or mining engineering maps; and performing supervised learning training on the 3D convolutional neural network model to enable it to automatically identify geological categories from physical attribute data.

[0013] In one preferred embodiment, the three-dimensional visualization is achieved through isosurface extraction algorithms (such as the moving cube algorithm) or volume rendering techniques, converting the voxel-format three-dimensional geological classification model into a three-dimensional surface model or volume data image.

[0014] As a preferred embodiment, the method further includes cross-validating the generated three-dimensional visualization model of the goaf using known information independent of the exploration process, such as borehole data or mining engineering drawings, to assess its accuracy.

[0015] A second aspect of the present invention provides a coal mine goaf identification exploration system, characterized in that it comprises: A multi-source data acquisition module is used to acquire at least two types of geophysical exploration data from the exploration area. The data preprocessing module is used to preprocess the geophysical exploration data. The cross-modal data fusion module is used to establish a three-dimensional grid covering the exploration area, map the pre-processed multiple physical attribute data to the nodes of the three-dimensional grid through spatial interpolation to form a high-dimensional physical attribute vector, and convert the high-dimensional physical attribute vector into a fusion feature vector in a unified feature space based on a preset cross-modal data fusion model. The intelligent analysis and inversion module is used to generate a three-dimensional geological classification model that characterizes the spatial location and attributes of the goaf through collaborative inversion and intelligent interpretation. The model generation and display module is used to perform three-dimensional visualization of the three-dimensional geological classification model and generate a three-dimensional visualization model of the goaf.

[0016] This invention provides a method for identifying coal mine goaf areas. It has the following beneficial effects: 1. This invention employs a deep autoencoder for data-driven feature fusion, mapping multi-source geophysical data with different physical units and dimensions to a unified low-dimensional feature space. This method can automatically learn and characterize the intrinsic correlations between different physical properties, overcoming the technical challenge of effectively fusing heterogeneous data using traditional methods, and providing a unified and robust data foundation for subsequent high-precision inversion and interpretation.

[0017] 2. This invention employs a collaborative inversion method based on structural coupling, placing multiple geophysical inversion problems within a unified optimization framework. By introducing structural constraints such as cross-gradient constraints, different physical property models are forced to maintain consistency in the structural change regions reflecting the boundaries of geological bodies. This achieves mutual constraints among multiple data sets, significantly reducing the inherent ambiguity of single-method inversion and improving the reliability and geological fit of the inversion results.

[0018] 3. This invention achieves automated voxel-level classification of three-dimensional multi-physical attribute models by introducing a three-dimensional convolutional neural network for model interpretation. This method replaces the traditional interpretation process that relies on human experience, avoids the uncertainty caused by subjective factors, and significantly improves the efficiency and objectivity of delineating goaf boundaries and classifying internal attributes, resulting in more refined and accurate exploration results. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0020] 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.

[0021] Example: Please see the appendix Figure 1 - Appendix Figure 2 This invention provides an exploration method for identifying coal mine goaf areas, comprising: This invention provides a survey method for identifying coal mine goaf areas. This method achieves refined detection of underground goaf targets by constructing an orderly workflow consisting of multiple technical steps.

[0022] The method includes: S1, multi-source collaborative data acquisition; S2, data preprocessing and feature extraction; S3, cross-modal data fusion; S4, collaborative inversion and intelligent interpretation; and S5, 3D model generation and verification.

[0023] In the specific implementation process, S1 is first executed, which involves deploying an integrated air-ground-well observation network in the area to be explored to acquire geophysical observation data with at least two different physical properties. These data form the basis for subsequent analysis, and their diversity ensures comprehensive constraints on the subsurface geological model.

[0024] Next, step S2 is executed to perform necessary denoising and correction processing on the acquired raw data to eliminate environmental and instrument interference and improve the data signal-to-noise ratio. Simultaneously, specific signal processing techniques are employed to identify and extract weak anomaly response features in the data related to the goaf.

[0025] Next, step S3 is executed. This step fuses preprocessed data from different physical exploration methods. The principle is that although the physical quantities observed by various methods are different, the target geological body is uniform. By aligning data from different sources in a unified spatial grid and constructing a common feature space, heterogeneous data are transformed into a unified format that can be collaboratively analyzed.

[0026] Based on data fusion, step S4 is executed. This step uses the unified feature space information established in step S3 as constraints to perform multi-physics collaborative inversion calculations, thereby obtaining a three-dimensional underground physical property model with consistent structural characteristics. Subsequently, an intelligent interpretation algorithm is used to analyze the model, automatically identifying and delineating the spatial range, shape, and internal filling state of the goaf.

[0027] Finally, step S5 is executed to visualize the goaf information interpreted in S4 in three dimensions, generating an intuitive, detailed structural model of the goaf. This model is then compared with known geological data or mining engineering drawings to verify the accuracy of the exploration results, resulting in a final assessment report.

[0028] To implement the above method, the present invention also provides a corresponding exploration system. (Refer to...) Figure 2 , Figure 2 This is a system structure block diagram according to an embodiment of the present invention. The system includes multiple interconnected functional modules.

[0029] Specifically, the system includes: The multi-source data acquisition module is used to execute S1 and acquire geophysical observation data of the area to be explored; The data preprocessing module, connected to the multivariate data acquisition module, is used to execute S2 to process the acquired data and extract features; The cross-modal data fusion module, connected to the data preprocessing module, is used to execute S3 and construct a unified feature space; The intelligent analysis and inversion module, connected to the cross-modal data fusion module, is used to execute S4 for collaborative inversion and intelligent interpretation. The model generation and display module, connected to the intelligent analysis and inversion module, is used to execute S5 to generate and display the final 3D model.

[0030] During system operation, field data acquired by the multi-data acquisition module is transmitted to the data preprocessing module. The processed data then flows to the cross-modal data fusion module for integration. The fused feature data is received and calculated by the intelligent analysis and inversion module. Finally, the model generation and display module receives the interpretation results, generates and presents the 3D exploration results to the user. These modules work collaboratively to form a complete and coherent technical implementation system.

[0031] This invention provides an exploration method for identifying coal mine goaf areas, comprising the following steps: S1, Multi-source data collaborative acquisition: Through an integrated air-ground-well observation network, acquire geophysical observation data of at least two different physical properties of the area to be explored; S2, Data Preprocessing and Feature Extraction: The collected raw data is denoised and corrected, and weak anomalous signal features reflecting geological anomalies are identified and extracted. S3, cross-modal data fusion: aligns multi-source, heterogeneous physical attribute data in a unified spatial grid and constructs a unified feature space; S4, Collaborative Inversion and Intelligent Interpretation: Under the constraints of a unified feature space, multi-physics collaborative inversion calculations are performed, and intelligent algorithms are used to automatically interpret the inversion results in order to identify goaf areas; S5, 3D Model Generation and Verification: The interpretation results are visualized in 3D to generate a fine structural model of the goaf, which is then compared and verified with known geological data.

[0032] The core of S1 lies in constructing an observation system that can provide multi-physical constraints. Single geophysical methods inherently suffer from multiple solutions in inversion interpretation, meaning different subsurface geological models may produce the same observational response. By collaboratively acquiring data reflecting different physical properties of the subsurface medium, the inherent correlations between these properties can be used to mutually constrain each other, thereby effectively converging the solution space and reducing interpretation uncertainty.

[0033] The specific implementation of this step may include the following sub-steps: Analyze geological targets and environmental conditions. Before commencing exploration, collect existing geological data, mining engineering maps, hydrogeological data, and physical property test data for the area to be explored. Based on this data, analyze and estimate the basic characteristics of the goaf, including its burial depth range, approximate spatial distribution, possible size, and potential filling material conditions (such as cavities, water infill, or collapsed gangue filling). Simultaneously, assess the topography, surface cover, and levels of human and industrial electromagnetic interference in the exploration area.

[0034] Design and implement method combinations. Based on the analysis results, select at least two geophysical methods that are sensitive to the target body and have complementary physical mechanisms for combination. The design of the method combination follows the principle of complementary advantages and cost-effectiveness.

[0035] For example, in one embodiment, for mined-out areas with shallow burial depths (e.g., less than 100 meters) and unknown filling conditions, an observation scheme combining high-density resistivity method and shallow seismic exploration (or micro-motion detection method) can be designed. High-density resistivity method is sensitive to electrical differences between the mined-out area and the surrounding rock and water-filled areas; while seismic exploration method is sensitive to changes in wave velocity and density of the rock and soil, effectively identifying low-velocity anomalies caused by mining or fracturing. The combination of the two methods can effectively distinguish between high-resistivity, low-velocity uncollapsed cavities, low-resistivity, low-velocity water-filled areas, and collapse zones with properties intermediate between the two.

[0036] In another embodiment, for deep-buried goaf areas suspected of having large-scale, continuous water filling, an observation scheme combining transient electromagnetic methods and gravity exploration methods can be designed. Transient electromagnetic methods have a high detection capability for deep, low-resistivity bodies and can effectively delineate the extent of water filling areas; while gravity exploration methods can reflect large-scale mass loss caused by mining activities, providing a basis for judging the macroscopic distribution of goaf areas.

[0037] The integrated air-ground-well observation system systematically combines observation methods from different platforms. For example, a large-area aerial scan can be conducted first using a UAV equipped with a magnetometer or electromagnetic system to quickly delineate key areas of anomaly development. Subsequently, a high-precision ground-based exploration network, such as high-density electrical resistivity tomography (EDT) lines, is deployed within the key areas. Finally, if existing boreholes exist on the site, they can be used for in-well geophysical exploration to obtain high-resolution physical property parameters at specific locations, which can then be used to calibrate and verify the results of ground and aerial surveys.

[0038] For the specific instrument operation and data acquisition procedures of each of the above-mentioned individual geophysical exploration methods, those skilled in the art can refer to relevant industry standards and technical specifications, as these are well-known technologies in the field and will not be elaborated upon here. All data acquired by these methods are recorded with precise three-dimensional spatial coordinates, providing a foundation for data fusion and spatial alignment in subsequent steps.

[0039] In the implementation of S1, in order to ensure that the collected multi-data can be effectively used for subsequent fusion and collaborative inversion, it is necessary to systematically deploy the measurement points of each exploration method in space and perform precise spatial location registration of all data.

[0040] The basic principle of survey network deployment is that its sampling density must meet the requirements for effectively resolving the expected size and depth of the target goaf. For example, for profiling exploration methods such as high-density resistivity methods or seismic exploration, survey lines are usually laid out perpendicular to the direction of the main geological structure of the area or the long axis of the goaf. The spacing between survey lines and the spacing between survey points are determined based on the estimated depth and lateral dimensions of the target body. For area exploration methods such as gravity, magnetic methods, or micro-motion arrays, regular or irregular grid-like survey points are laid out within the exploration area, and the density of survey points is determined according to the required resolution and coverage.

[0041] Spatial registration of acquired data is a crucial prerequisite for subsequent cross-modal data fusion. This process ensures that all observation data from different methods and platforms are assigned three-dimensional coordinates in the same unified coordinate system.

[0042] In practice, a unified coordinate system covering the entire work area is established before the exploration work begins. This coordinate system can be a national standard geodetic coordinate system or an independent engineering coordinate system. During data acquisition, high-precision differential global positioning systems and other measuring equipment are used to accurately determine the three-dimensional coordinates of each ground measuring point (such as electrodes, detectors, and gravity points). For aerial exploration, the airborne inertial measurement unit (IMU) works in conjunction with GPS to record the precise position of the flight platform in real time. For well drilling, the spatial positions of sensors at various depths are calculated based on the wellhead coordinates and inclination data.

[0043] In S2, the collected raw data is corrected and denoised to eliminate systematic errors and random interference caused by non-geological target factors, thereby improving the signal-to-noise ratio of the data and providing a high-quality data foundation for subsequent weak anomaly signal extraction and data fusion.

[0044] Data correction processing varies depending on the geophysical method used. In one embodiment, if the acquired data includes gravity or magnetic data, the correction process includes normal field correction, topographic correction, and mesosphere correction to eliminate the effects of the standard Earth ellipsoid model, topographic relief, and inhomogeneities in material density between the surface and the reference surface. For magnetic data, diurnal variation correction is also required to account for changes in the geomagnetic field over time. In another embodiment, if the acquired data is high-density resistivity resistivity data, the correction process primarily involves topographic correction to eliminate apparent resistivity distortion caused by surface undulations.

[0045] Data denoising primarily targets power frequency interference, instrument system noise, and environmental random noise. In practice, digital filtering methods can be used to suppress these noises. For example, notch filters can be applied to power frequency interference at 50Hz or its harmonics. Low-pass or band-pass filters can be used to preserve the effective signal bandwidth for high-frequency random noise. For seismic data, methods such as FK filtering can also be used to suppress coherent noise.

[0046] For the specific mathematical implementation of the above correction and filtering algorithms, those skilled in the art can refer to relevant geophysical data processing textbooks and software manuals, which are well-known technologies in this field and will not be elaborated here.

[0047] During the implementation of S2, after conventional correction and denoising, weak anomalous signals may still exist in the data, caused by deep or small goaf areas and superimposed on a strong background field. To achieve precise detection of these targets, further identification and extraction of these weak anomalous signals are required.

[0048] In one embodiment of the invention, wavelet transform is used to perform multi-scale analysis of the signal to identify and extract the weak anomalous signal. For one-dimensional geophysical profile data... Its continuous wavelet transform Defined as: ; in: Represents the position along the survey line Changing geophysical observation signals; The scaling factor reflects the frequency scale of the analysis; , where is the translation factor, representing the center position of the wavelet function on the signal; For the mother wavelet function; It is the complex conjugate of the mother wavelet function.

[0049] Wavelet transform can reveal the local characteristics of signals at different scales and locations. Geological anomalies (such as the boundaries of goaf areas) often cause local singularities in signals, which manifest as the propagation of modulus maxima along the scale direction in the wavelet transform domain. By detecting and tracing these modulus maxima, the location of weak anomalous signals on a spatial profile can be determined. Based on this, wavelet coefficients at specific scales can be thresholded or reconstructed, thereby enhancing weak anomalous signals or separating them from the background field.

[0050] In another embodiment, this step can also employ a deep learning-based signal separation algorithm. For example, a denoising autoencoder network can be constructed and trained. This network takes the original signal segment containing weak anomalies as input and learns to reconstruct the background signal without anomalies through an encoding-decoding process. Subtracting the network-reconstructed background signal from the original input signal yields the extracted weak anomaly signal.

[0051] The output of this step is a set of data with a significantly improved signal-to-noise ratio, in which the geophysical response characteristics related to the goaf are effectively highlighted, providing a foundation for subsequent cross-modal data fusion and collaborative inversion.

[0052] The core of S3 lies in solving the problem of effectively integrating the multi-source, heterogeneous geophysical data obtained in S2. Its theoretical basis is the common low-dimensional manifold hypothesis, which posits that different physical observations (such as resistivity, wave velocity, and density) of the same geological body (e.g., surrounding rock, mined-out area, caving zone) are high-dimensional representations of the body's intrinsic properties in different physical dimensions, all distributed along a lower-dimensional underlying manifold structure. The goal of this step is to find this common low-dimensional manifold and construct a unified feature space capable of characterizing it.

[0053] The specific implementation of this step may include the following sub-steps: Data Spatial Alignment and Meshization. A three-dimensional Cartesian grid covering the entire exploration area is established. All S2-processed geophysical data are mapped to the nodes of this three-dimensional grid using spatial interpolation algorithms (such as Kriging interpolation or radial basis function interpolation). Above. After this step, each node in the 3D mesh corresponds to a high-dimensional physical property vector composed of multiple physical property values. : ; in, This represents the number of types of physical properties. For the first The value of a physical property at this node.

[0054] Constructing a unified feature space. A preferred implementation is to use a deep autoencoder network for feature fusion. A deep autoencoder consists of an encoder and a decoder. The encoder transforms the high-dimensional physical property vector through a nonlinear transformation composed of multiple neural network layers (e.g., fully connected layers and activation function layers). Compressed into a lower-dimensional fusion feature vector This process can be represented as: ; The decoder part then receives this low-dimensional fused feature vector. And it attempts to reconstruct the original high-dimensional physical property vector through a symmetric nonlinear transformation, thus obtaining the reconstructed vector. This process can be represented as: ; in: A function representing the encoder network; A function representing the decoder network; This is the desired low-dimensional fused feature vector in the unified feature space.

[0055] The training objective of the network is to minimize the original input vector. With reconstructed vector The difference between them, i.e., minimizing the reconstruction error. This error is typically calculated using the mean squared error loss function. To measure: ; in, This represents the total number of grid nodes participating in the training.

[0056] The network parameters are continuously adjusted using the backpropagation algorithm until the loss function is reached. Convergence. After training, the encoder part of the network... It can be used as a fixed feature converter. It converts the high-dimensional physical attribute vectors of all grid nodes in the exploration area... By inputting this encoder, a set of corresponding low-dimensional fused feature vectors that can characterize the comprehensive properties of geological bodies in a unified feature space can be obtained. This fused feature vector will serve as input for the next step of collaborative inversion and intelligent interpretation.

[0057] The theoretical basis for this step is the common low-dimensional manifold hypothesis. Its physical significance is that for geological units in the same underground spatial location, their different physical properties (such as resistivity, seismic wave velocity, density, etc.) are not independent of each other, but are jointly determined by intrinsic factors such as the rock and mineral composition, pore structure, fluid saturation, and physical state of the geological unit.

[0058] Therefore, when multiple physical properties are used as coordinate axes to construct a high-dimensional feature space, data points representing specific geological types (such as intact surrounding rock, collapsed and fractured rock mass, or water-filled areas) are not randomly scattered throughout the space. Instead, they cluster and are distributed on an embedded geometric structure with a dimension much lower than that of the high-dimensional space. This structure is called a common low-dimensional manifold.

[0059] The fundamental purpose of S3 is to learn and represent this common low-dimensional manifold through a data-driven approach. By mapping a high-dimensional vector composed of multiple physical properties to a point on this low-dimensional manifold, a fused feature vector can be obtained. This vector can more essentially and compactly describe the comprehensive geological and geophysical characteristics of this spatial location, thereby eliminating redundant information and providing a more effective input for subsequent collaborative inversion and intelligent interpretation.

[0060] In S4, a structurally coupled collaborative inversion is performed first. The purpose of this step is to overcome the ambiguity of single-method inversion by jointly optimizing multiple geophysical models to ensure that the final physical property models (such as resistivity models and velocity models) are consistent in spatial structure.

[0061] This collaborative inversion is achieved by constructing and minimizing a joint objective function. This is achieved by simultaneously inverting two physical property models. In one embodiment, it is assumed that two physical property models are simultaneously inverted. and Then the joint objective function can be expressed as: ; in: These represent two discretized three-dimensional physical property models to be inverted (e.g., For resistivity model, (for velocity model) This is the data fitting term, used to measure the difference between the model's forward response and the actual observed data. Its form is typically: ,in It is observation data. It is a forward computation operator. It is a data weighted matrix; This is a model regularization term used to impose prior constraints on the solution to ensure the stability and smoothness of the inversion results. Its typical form is... ,in It is a model weighting matrix. In one embodiment, the weighting matrix may be a gradient or Laplacian operator matrix, used to apply model smoothing constraints. This is the structural coupling term, and it is the core of this cooperative inversion. This term is used to force... and They tend to be structurally consistent; , , This is a regularization parameter used to balance the weights of each objective function term.

[0062] Structural coupling terms In a preferred embodiment, a cross-gradient function is used. For the model and gradient at any point in space and Its gradient vector Defined as: ; The magnitude of the cross-gradient vector is zero or close to zero when the gradient directions of the two models are parallel or antiparallel. This characteristic corresponds precisely to the boundary of a geological body, where different physical properties often undergo abrupt changes simultaneously, and their gradient directions are highly correlated. Therefore, the structural coupling term is defined as the integral or summation of the cross-gradient magnitudes over the entire model space: ; By minimizing the joint objective function The inversion process will search for a set of models. , This ensures that while fitting the observed data, the internal structural boundaries of these structures remain as consistent as possible. The optimization problem can be solved using iterative algorithms such as the nonlinear conjugate gradient method; their specific implementations are well-known techniques in this field and will not be elaborated upon here.

[0063] Ultimately, this step outputs a set of three-dimensional physical property models with common structural features, providing highly reliable input for subsequent intelligent interpretation.

[0064] Following collaborative inversion, traditional model interpretation relies on human experience, resulting in low efficiency and high subjectivity. To achieve automatic delineation and attribute classification of goaf boundaries, this invention further employs an intelligent algorithm in S4 to interpret the three-dimensional multi-physical attribute model obtained through collaborative inversion.

[0065] A preferred implementation of this step is to use a three-dimensional convolutional neural network for voxel-level classification. 3D-CNN can directly process three-dimensional data volumes and learn and extract complex spatial features from them, making it suitable for recognizing the three-dimensional morphology of geological bodies.

[0066] The intelligent interpretation process may include the following sub-steps: Construction of Training Samples. A sample set for training the 3D-CNN is constructed. This sample set consists of data-label pairs. "Data" refers to 3D data blocks extracted from known geological models or areas with borehole verification. Each data block is a small 3D volume, with each voxel containing multiple physical property values ​​obtained through co-inversion (e.g., resistivity and velocity values). "Label" is the actual geological category corresponding to the central voxel of the data block, such as "surrounding rock," "mined-out area," "collapse zone," or "water-filled area." This label information is derived from existing geological maps, mining engineering maps, or borehole core logging data.

[0067] Training the network model. The 3D-CNN model is trained using a constructed sample set through supervised learning. During training, the network learns to extract features that distinguish different geological categories from the input 3D multi-physical attribute data blocks by optimizing the loss function. After training, a classification model is obtained that can automatically determine the geological category based on the input physical attribute data. For the specific construction and training process of the 3D-CNN network, those skilled in the art can refer to relevant deep learning technologies, which are well-known in the field and will not be elaborated upon here.

[0068] Automatic model interpretation. The trained 3D-CNN model is applied to the three-dimensional multi-physical attribute data volume of the entire exploration area. The model traverses the entire data volume in a sliding window manner, analyzing each voxel and its neighborhood data, and outputting the probability of the geological category to which the voxel belongs. By selecting the category with the highest probability as the final classification result for that voxel, a three-dimensional geological classification model is generated. This model automatically marks the spatial location, morphological boundaries, and internal filling status of the goaf at the voxel level.

[0069] This step aims to convert the three-dimensional geological classification model generated in S4, which is expressed in voxel form, into a three-dimensional visualization model that can be intuitively displayed and analyzed, so as to clearly present the location, shape, scale and relationship of the goaf in the underground space with the surrounding environment.

[0070] In a preferred embodiment, the visualization model is primarily generated using an isosurface extraction algorithm. Specifically, a moving cube algorithm can be applied to the 3D geological classification model. This algorithm considers voxels labeled "mining area," "collapse zone," or "water-filled area" in the classification results as target regions. By examining each cube cell in the 3D mesh one by one, it determines whether the geological body boundary passes through the cell based on the geological category of its eight vertices. If the boundary passes through, one or more triangular facets are generated within the cell to approximate the local boundary. Connecting all the triangular facets generated in the entire model space ultimately forms a closed 3D surface model representing the outline of target geological bodies such as mining areas.

[0071] In another embodiment, volume rendering technology can also be used to show the gradual changes in properties within the goaf (such as the degree of collapse and fragmentation, and water saturation). This technology does not generate explicit geometric surfaces, but rather generates a semi-transparent 3D image that reflects the changes in the internal structure and properties of the model by assigning specific colors and opacities to each voxel in the 3D data volume based on its physical property values ​​or classification labels, and then performing ray projection integration along the line of sight.

[0072] For the specific programming implementation of the moving cube algorithm and volume drawing algorithm, those skilled in the art can refer to relevant computer graphics technologies, which are well-known technologies in this field and will not be elaborated here.

[0073] The final output 3D visualization model can be loaded into a professional software platform for interactive operations such as rotation, scaling, and sectioning at any angle, providing an intuitive basis for subsequent engineering decisions.

[0074] To verify the accuracy and reliability of the three-dimensional model of the goaf generated by the method of this invention, step S5 also includes a step of cross-validation of the model. This step evaluates the accuracy of the model by comparing the model interpretation results with known information independent of the detection process of this invention.

[0075] In one embodiment, this cross-validation utilizes existing borehole data or data specifically constructed for validation purposes. The process imports the borehole's 3D spatial trajectory and its core logging geological log into the same coordinate system as the 3D visualization model. Subsequently, the two are compared spatially. For example, if the borehole's geological log records a "cavity" or "fracture zone" at a certain depth, the model voxels at the corresponding depth are checked to see if they are classified as "mined-out area" or "collapse zone" by the intelligent interpretation algorithm. By statistically analyzing the match rate between the model's classification results and the actual borehole findings, the model's local accuracy can be quantitatively evaluated.

[0076] In another embodiment, this cross-validation can also utilize historical data such as mining engineering maps or geological maps of the mine. After digitizing and spatially registering these drawings, they are overlaid on the generated 3D model. By observing the degree of spatial overlap between the goaf area marked on the drawings and the goaf boundary interpreted by the model, the accuracy of the macroscopic location, shape, and scale of the model interpretation results can be qualitatively or semi-quantitatively assessed.

[0077] This verification step provides an objective basis for the reliability of the final results and can provide feedback for the adjustment and optimization of inversion and interpretation parameters.

[0078] 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 identifying coal mine goaf areas, characterized in that, Includes the following steps: S1: Obtain at least two types of geophysical exploration data from the exploration area; S2: Perform data preprocessing on the geophysical exploration data; S3: Establish a three-dimensional grid covering the exploration area, and map the preprocessed physical attribute data to the nodes of the three-dimensional grid through spatial interpolation to form a high-dimensional physical attribute vector; S4: Based on the preset cross-modal data fusion model, the high-dimensional physical attribute vector is converted into a fusion feature vector in a unified feature space; S5: Through collaborative inversion and intelligent interpretation, a three-dimensional geological classification model representing the spatial location and attributes of the goaf is generated; S6: Perform three-dimensional visualization on the three-dimensional geological classification model to generate a three-dimensional visualization model of the goaf.

2. The method according to claim 1, characterized in that, The cross-modal data fusion model is a deep autoencoder network; the step of converting the high-dimensional physical attribute vector into a fusion feature vector in a unified feature space includes: using the high-dimensional physical attribute vector as input to the deep autoencoder network, and obtaining the fusion feature vector through its encoder.

3. The method according to claim 1, characterized in that, The collaborative inversion is achieved by constructing and minimizing a joint objective function, which includes a data fitting term, a model regularization term, and a structural coupling term.

4. The method according to claim 3, characterized in that, The structural coupling term is a cross-gradient function used to constrain the spatial structural consistency of models with different physical properties.

5. The method according to claim 1, characterized in that, The intelligent interpretation is achieved through a pre-trained intelligent interpretation model, which is used to perform voxel-level classification on the three-dimensional multi-physical attribute model obtained by collaborative inversion, thereby obtaining the three-dimensional geological classification model.

6. The method according to claim 5, characterized in that, The intelligent interpretation model is a three-dimensional convolutional neural network model.

7. The method according to claim 6, characterized in that, The training process of the three-dimensional convolutional neural network model includes: constructing a training sample set using three-dimensional data blocks containing geological labels, and performing supervised learning training on the three-dimensional convolutional neural network model.

8. The method according to claim 1, characterized in that, The 3D visualization of the 3D geological classification model is achieved through isosurface extraction algorithms or volume rendering techniques.

9. The method according to claim 1, characterized in that, The method further includes: The three-dimensional visualization model of the goaf is cross-validated using known information such as borehole data or mining engineering drawings.

10. A coal mine goaf identification exploration system, and a coal mine goaf identification exploration method according to any one of claims 1-9, characterized in that, include: A multi-source data acquisition module is used to acquire at least two types of geophysical exploration data from the exploration area. The data preprocessing module is used to preprocess the geophysical exploration data. The cross-modal data fusion module is used to establish a three-dimensional grid covering the exploration area, map the pre-processed multiple physical attribute data to the nodes of the three-dimensional grid through spatial interpolation to form a high-dimensional physical attribute vector, and convert the high-dimensional physical attribute vector into a fusion feature vector in a unified feature space based on a preset cross-modal data fusion model. The intelligent analysis and inversion module is used to generate a three-dimensional geological classification model that characterizes the spatial location and attributes of the goaf through collaborative inversion and intelligent interpretation. The model generation and display module is used to perform three-dimensional visualization of the three-dimensional geological classification model and generate a three-dimensional visualization model of the goaf.

Images