High-density earthquake and electrical method combined detection and identification method for goaf of strip mine

Abstract

The invention relates to the technical field of strip mine goaf geophysical detection, and discloses a strip mine goaf high-density earthquake and electrical method combined detection and identification method, which comprises the following steps: obtaining strip mine target area high-density earthquake and electrical method observation data; preprocessing the acquired data; performing deep feature extraction on the preprocessed data, and fusing seismic and electrical method features by using cross-modal attention to obtain fused features; based on the fusion features, a loss function containing data fitting, a physical equation and structural coupling constraint is used for guiding deep learning joint inversion, and an underground three-dimensional speed and resistivity model is obtained; and on the basis of the inversion model, identifying and parameterizing the representation goaf by using an intelligent identification network to obtain an identification result and parameterization information of the representation goaf. Through intelligent fusion and fine inversion, the goaf detection precision and the automation and parameterization characterization capability are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of geophysical detection of open-pit mine goafs, and in particular to a method for combined high-density seismic and electrical detection and identification of open-pit mine goafs. Background Art

[0002] During the mining process of many large open-pit mines in my country, underground goafs with unclear size, depth and filling conditions are often formed below the surface, posing a serious threat to mining safety and deep resource extraction. Therefore, timely and accurate detection and detailed characterization of goafs are crucial.

[0003] Currently, the detection of hidden goaf in open-pit mines is primarily done through geophysical exploration supplemented by drilling verification. As mining deepens and operating space becomes more limited, the industry is leaning towards high-density observations to collect richer data to improve detection accuracy. Despite the combined use of multiple geophysical methods, such as seismic and electrical methods, existing technologies often process and interpret data from each method independently, failing to achieve deep fusion and collaborative inversion of multi-source data at the information level. This limits the overall effectiveness of integrated geophysical exploration techniques.

[0004] However, existing technologies have shortcomings in the detailed detection of complex open-pit mine goafs. First, at the level of multi-source data feature extraction and fusion, it is difficult to fully explore the complex features related to goafs, and the information fusion efficiency is low, resulting in an incomplete understanding of their comprehensive geophysical anomalies. Secondly, the inversion imaging link lacks effective physical process guidance and structural coupling constraints across physical parameter models, which limits the accuracy and geological conformity of the reconstructed underground structure images, affecting the precise characterization of goafs. Furthermore, the subsequent goaf identification and parameterized description process relies on manual interpretation, which is inefficient and highly subjective, making it difficult to achieve a comprehensive and automated quantitative description, restricting the application value of the detection results.

[0005] Therefore, the present invention proposes a method for joint detection and identification of open-pit mine goafs by high-density seismic and electrical methods to address the deficiencies of the prior art. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for the joint detection and identification of open-pit mine goafs using high-density seismic and electrical methods, which solves the problems of insufficient detection accuracy and comprehensive cognitive depth in the existing technology due to insufficient data fusion and inversion constraints, and lack of automated and fine characterization of goafs.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for combined high-density seismic and electrical detection and identification of open-pit mine goafs, comprising the following steps: S1. Acquire high-density seismic observation data and high-density electrical observation data for the target area of ​​the open-pit mine; S2. performing data preprocessing on the acquired high-density seismic observation data and high-density electrical observation data to obtain preprocessed seismic data and preprocessed electrical data; S3. performing deep feature extraction on the preprocessed seismic data and the preprocessed electrical data, and fusing the extracted seismic features and electrical features using a cross-modal attention mechanism to obtain fused features; S4. Based on the fusion features, a deep learning joint inversion network is used to perform joint inversion to obtain a three-dimensional velocity model and a three-dimensional resistivity model of the underground. The training process of the joint inversion network adopts a comprehensive loss function including a data fitting term, a physical equation constraint term, and a structural coupling constraint term to guide physical information. S5. Based on the joint inversion, a three-dimensional velocity model and a three-dimensional resistivity model of the underground are obtained, and an intelligent recognition network is used to identify and parameterize the goaf to obtain the recognition result and parameterization information of the goaf.

[0008] Preferably, the step S1 includes: Develop high-density seismic data acquisition plans and high-density electrical data acquisition plans based on the geological conditions and exploration tasks of the target area of ​​the open-pit mine. The high-density seismic data acquisition plans and high-density electrical data acquisition plans include observation geometry parameters and data acquisition operation parameters for high-density coverage. According to the developed high-density seismic data acquisition plan, seismic data acquisition operations are carried out in the target area of ​​the open-pit mine to obtain high-density seismic observation data; According to the formulated high-density electrical data acquisition plan, electrical data acquisition operations are carried out in the target area of ​​the open-pit mine to obtain high-density electrical observation data.

[0009] Preferably, the step S2 includes: For the acquired high-density seismic observation data, field record editing, observation channel geometry information assignment, noise suppression, static correction and amplitude compensation processing are performed to obtain the pre-processed seismic data; For the acquired high-density electrical observation data, invalid data point elimination, zero point correction, terrain correction and electrode grounding condition check are performed to obtain the pre-processed electrical data.

[0010] Preferably, in step S3, the step of performing depth feature extraction on the preprocessed seismic data and the preprocessed electrical data includes: Inputting the preprocessed seismic data into a seismic feature extraction deep neural network to extract deep abstract features of the seismic data; The pre-processed electrical method data is input into an electrical method feature extraction deep neural network to extract deep abstract features of the electrical method data.

[0011] Preferably, in step S3, the step of fusing the extracted seismic features and electrical features using a cross-modal attention mechanism to obtain fused features comprises: Inputting the extracted deep abstract features of the seismic data and the deep abstract features of the electrical data into a cross-modal attention module based on a transformer encoder structure; Inside the cross-modal attention module, the query matrix Q is generated by projecting the deep abstract features of one modality, and the key matrix K and value matrix V are generated by projecting the deep abstract features of the other modality, or vice versa, to perform cross-modal interaction calculations; Using the generated query matrix Q, the key matrix K, and the value matrix V, the attention weighted output is calculated using the following formula to capture and quantify the correlation between the two modal features: Where K T is the transpose of the key matrix K; d k is the dimension of the key vector in the key matrix k; softmax is the normalized exponential function; V is the value matrix; Based on the attention weighted output calculated by applying the attention mechanism once or multiple times, a fusion feature is generated, where the fusion feature integrates complementary information of the seismic data and the electrical data.

[0012] Preferably, in step S4, the step of using a deep learning joint inversion network to perform joint inversion based on the fusion features to obtain a three-dimensional velocity model and a three-dimensional resistivity model of the underground includes: Inputting the fused features into a pre-configured deep learning joint inversion network; The deep learning joint inversion network reconstructs the three-dimensional seismic wave velocity distribution and the three-dimensional resistivity distribution of the target area of ​​the open-pit mine in parallel or jointly based on the fusion features; The three-dimensional seismic wave velocity distribution reconstructed by the deep learning joint inversion network is output as a three-dimensional velocity model, and the three-dimensional resistivity distribution is output as a three-dimensional resistivity model.

[0013] Preferably, in step S4, the training goal of the deep learning joint inversion network is to minimize a comprehensive loss function L total , the comprehensive loss function L total Defined as: L total =λ data L data +λ phyL phy +λ couple L couple ; Where, L data is the data fitting term; L phy is the constraint term of the physical equation; L couple is the structural coupling constraint term; data ,λ phy ,λ couple are the non-negative weight coefficients of the data fitting term, the physical equation constraint term and the structural coupling constraint term respectively.

[0014] Preferably, in step S5, the step of obtaining the underground three-dimensional velocity model and the three-dimensional resistivity model based on the joint inversion and using an intelligent recognition network to identify and parameterize the goaf includes: providing the three-dimensional velocity model and the three-dimensional resistivity model as input to a pre-trained intelligent recognition network; The intelligent recognition network processes the input to automatically identify and segment the three-dimensional spatial area of ​​the underground goaf; for each goaf identified by the intelligent recognition network, its geometric parameters and internal average physical properties are calculated to form a parametric representation of the goaf.

[0015] Preferably, the identification result and parameterized characterization information of the goaf obtained in step S5 include: The identification result is a digital representation of the three-dimensional spatial position and shape of each goaf; The parameterized characterization information is a set of quantitative descriptors D determined for each identified goaf g. g , the set D g It is composed in the following ways: {ID g ,Geometry g ,Physics g}; Where, ID g It is the unique identifier of the goaf; Geometry g is a set of parameters that characterize the geometric characteristics of the goaf; Physics g It is a set of parameters that characterize the average physical properties inside the goaf.

[0016] In summary, the present invention includes at least one of the following beneficial technical effects: 1. This invention fuses high-density seismic and electrical data features through deep learning and a cross-modal attention mechanism, improving the ability to detect anomalies in open-pit mine goafs. Combined with physical information-guided joint inversion and intelligent recognition, it accurately defines the location and morphology of goafs, overcoming the shortcomings of traditional methods in feature extraction, information fusion, and inversion constraints, and improving recognition accuracy and reliability.

[0017] 2. Based on a three-dimensional velocity and resistivity model, this invention uses an intelligent recognition network to provide a detailed parametric characterization of goaf areas, including their geometry and internal average physical properties. This multi-dimensional quantitative description deepens understanding of goaf structure, filling status, and risks, providing comprehensive data support for refined open-pit mine management and disaster prevention, and enhancing scientific decision-making.

[0018] 3. This invention incorporates deep learning technology throughout the entire process of high-density geophysical data processing, automating feature extraction, information fusion, joint inversion, and intelligent goaf identification and parameterization. This intelligent strategy improves detection efficiency and objectivity, reduces manual reliance, and meets the rapid response and massive data processing requirements of large-scale open-pit mine safety monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0020] The following is combined with Figure 1 , the present invention is described in further detail.

[0021] An embodiment of the present invention provides a method for joint detection and identification of open-pit mine goafs using high-density seismic and electrical methods, comprising the following steps: S1. Acquire high-density seismic observation data and high-density electrical observation data for the target area of ​​the open-pit mine; S2. Preprocessing the acquired high-density seismic observation data and high-density electrical observation data to obtain preprocessed seismic data and preprocessed electrical data; S3. Perform deep feature extraction on the preprocessed seismic data and the preprocessed electrical data, and fuse the extracted seismic features and electrical features using a cross-modal attention mechanism to obtain fused features; S4. Based on the fusion features, a deep learning joint inversion network is used to perform joint inversion to obtain the underground 3D velocity model and 3D resistivity model. The training process of the joint inversion network uses a comprehensive loss function containing data fitting terms, physical equation constraints, and structural coupling constraints to guide physical information. S5. Based on the joint inversion, the underground three-dimensional velocity model and three-dimensional resistivity model are obtained, and the goaf is identified and parameterized using an intelligent recognition network to obtain the identification results and parameterized characterization information of the goaf.

[0022] Regarding step S1, in this embodiment, as the cornerstone of the overall detection and identification process, its core lies in obtaining original observation data that can provide sufficient information support for subsequent high-precision geophysical inversion and intelligent goaf identification.

[0023] First, based on the geological conditions and exploration tasks of the target open-pit mine area, a high-density seismic data acquisition plan and a high-density electrical data acquisition plan are developed. This planning stage begins with a detailed survey and analysis of the target area. This analysis not only covers the regional geological and structural background, lithologic distribution characteristics, and hydrogeological conditions, but also fully considers the expected occurrence state of the goaf, such as its spatial location, approximate burial depth range, estimated size and morphology, and potential filling characteristics. Based on this comprehensive assessment, and closely centered around the specific exploration task requirements, such as preliminarily delineating the approximate scope of the goaf or characterizing its three-dimensional spatial structure, the present invention will construct targeted high-density seismic data acquisition plans and high-density electrical data acquisition plans.

[0024] The core content of high-density seismic data acquisition scheme and high-density electrical data acquisition scheme lies in establishing observation geometry parameters and data acquisition operation parameters that can achieve high-density coverage of the target area.

[0025] For high-density seismic data acquisition schemes, the observation geometry parameters are designed to achieve sufficient and uniform seismic wavefield illumination of underground targets, particularly potential goaf areas, while achieving high coverage. These parameters preferably specify the layout of seismic lines, such as densely packed two-dimensional profiles or three-dimensional observation grids. Furthermore, compared to conventional exploration, closer shot and receiver spacing is employed to significantly increase spatial sampling density. Furthermore, the minimum and maximum offsets are optimized to ensure good detection capabilities for targets at varying depths. Data acquisition operational parameters further refine the source type appropriate for the target geology (e.g., impact sources, vibroseis sources, or small, environmentally friendly sources for varying surface conditions and depths), the selection of high-sensitivity, broadband digital geophones, and the sampling rate and record length required to fully record effective wavefield information and meet the high-frequency requirements of subsequent processing. This scheme is designed to lay a solid foundation for the subsequent extraction of accurate seismic velocity information through refined spatial sampling and high-quality wavefield recording.

[0026] Correspondingly, for high-density electrical data acquisition schemes, the setting of its observation geometry parameters is also guided by the acquisition of high-density, high-resolution electrical structure information. Preferably, these parameters specify the specific direction and coverage of the electrical survey line, as well as the choice of electrode arrangement. For example, the Wenner device, Schellenberg device, dipole-dipole device commonly used in high-density resistivity imaging, or optimized arrangement combinations for specific targets can be used, and the electrode spacing and measurement point spacing can be accurately set to obtain a detailed description of the electrical differences at different depths and lateral ranges in the underground. Its data acquisition operation parameters focus on ensuring the stability of power supply parameters and the accuracy of power supply current, and implement strict data acquisition process quality control, such as real-time monitoring and control of electrode ground resistance, and checking the consistency of repeated observation data, so as to ensure that the final high-density electrical observation data is highly reliable and accurate.

[0027] Once the detailed data acquisition plan is developed, field seismic data acquisition will be carefully organized and implemented within the target area of ​​the open-pit mine using specialized seismic exploration instrumentation. By stimulating an artificial seismic source and receiving seismic waves reflected or refracted by the subsurface medium, high-density seismic observation data, which contains information on the elastic parameters of the subsurface medium (primarily reflecting differences in velocity structure), will be acquired. This data is typically presented as a multi-component, multi-channel raw time series record.

[0028] At the same time, we will strictly follow the established high-density electrical data acquisition plan and organize and implement field electrical data acquisition operations within the target area of ​​the open-pit mine using a professional electrical exploration instrument system. By supplying a steady current into the ground and measuring the resulting potential difference distribution at the surface, we can obtain a large amount of high-density electrical observation data that can reflect the differences in conductivity (i.e., resistivity) of the underground medium. This data is usually presented in the form of apparent resistivity values.

[0029] Through the above-mentioned systematic program design and rigorous field operation implementation, this step S1 ultimately obtains high-density seismic observation data and high-density electrical observation data for the target area of ​​the open-pit mine.

[0030] For step S2, in this embodiment, the high-density seismic observation data and high-density electrical observation data obtained in the above steps are preprocessed to obtain preprocessed seismic data and preprocessed electrical data, respectively. The specific technical implementation methods and considerations are described below. This step is a key link connecting the acquisition of original data with subsequent in-depth analysis. Its core goal is to effectively eliminate or minimize various types of noise and interference in the original observation data through a series of systematic and refined data processing processes, accurately correct the systematic errors introduced by observation conditions and environmental factors, and strive to restore and enhance the effective signals reflecting the underground geological structure. Only in this way can high-quality and high-reliability input data be provided for subsequent deep feature extraction, joint inversion and intelligent identification of goaf areas.

[0031] For the acquired high-density seismic observation data, this method will sequentially perform key processing including field record editing, observation channel geometry information assignment, multi-strategy noise suppression, refined static correction and comprehensive amplitude compensation, and finally obtain the preprocessed seismic data.

[0032] Field record editing is intended to preliminarily screen the original records, eliminate invalid traces, waste shots, or severely distorted traces, and ensure that subsequent processing is based on valid data.

[0033] The assignment of observation channel geometric information requires that the precise three-dimensional coordinates and elevation information of the shot point and receiver point be loaded into the header of each recorded channel, which is crucial for subsequent static correction, velocity analysis and imaging positioning.

[0034] Noise suppression processing aims to eliminate random noise and regular interference (such as surface waves and electromagnetic interference). Multi-domain filtering techniques (such as frequency domain and FK domain filtering, wavelet transform or deep learning denoising) are preferably used to significantly improve the signal-to-noise ratio and highlight the effective reflection signal.

[0035] The purpose of static correction is to eliminate travel time differences caused by near-surface velocity changes and terrain undulations. It usually includes elevation static correction, weathering layer static correction, and possibly residual static correction to ensure accurate alignment of the phase axis.

[0036] Amplitude compensation processing aims to correct the effects of spherical diffusion, dielectric absorption attenuation, and differences in coupling conditions on the amplitude, restore the relative relationship of the amplitudes, or perform true amplitude recovery. Common methods include exponential gain, automatic gain control, or Q compensation.

[0037] After the above processing, the final preprocessed seismic data has a high signal-to-noise ratio and fidelity, providing a high-quality foundation for subsequent analysis.

[0038] For the acquired high-density electrical observation data, this method will perform preprocessing including invalid data point removal, zero point correction, terrain correction and electrode grounding condition inspection to obtain preprocessed electrical data.

[0039] Invalid data points are eliminated based on the quality control parameters during acquisition (such as ground resistance, repeatability error, and statistical deviation), and apparent resistivity observation points that are obviously wrong or do not conform to physical laws are eliminated.

[0040] Zero point correction or drift correction is intended to eliminate instrument zero point drift or slowly changing background interference through baseline correction or trend surface fitting.

[0041] Terrain correction is crucial, especially in areas with high topography. Using an accurate digital elevation model, numerical simulations are performed to calculate and deduct the effect of topography on the apparent resistivity response, resulting in a more realistic distribution of underground electrical properties.

[0042] Electrode grounding condition checks assess the grounding status of each electrode by analyzing contact resistance data or comparing forward and reverse power supply results. Points with poor grounding can be weighted down or eliminated in subsequent inversions.

[0043] After the above processing, the pre-processed electrical data finally obtained can more accurately reflect the real electrical distribution characteristics of the underground medium.

[0044] In summary, step S2 significantly improves data quality and reliability by systematically preprocessing the original high-density seismic and electrical observation data.

[0045] In step S3, in this embodiment, deep feature extraction is performed on the preprocessed seismic data and the preprocessed electrical data, and the extracted seismic and electrical features are fused using a cross-modal attention mechanism to obtain fused features. This step is dedicated to automatically mining and refining the most recognizable and indicative deep abstract information of the target geological body (such as the open-pit mine goaf) from multi-source heterogeneous geophysical data through advanced deep learning models. It also strengthens the synergy and complementarity between these heterogeneous features through an intelligent feature fusion strategy, thereby generating a more comprehensive and robust feature representation.

[0046] This step begins with deep feature extraction on both preprocessed seismic and electrical data. Specifically, the preprocessed seismic data is fed into a purpose-built deep neural network for seismic feature extraction. Through its deep structure, this network automatically learns from the input seismic data and extracts deep, abstract features that characterize variations in elastic parameters and structural discontinuities in the subsurface. These features are more capable of capturing complex patterns within the data than traditional, manually selected geophysical attributes.

[0047] In parallel, the preprocessed electrical data is fed into a purpose-built deep neural network for electrical feature extraction. This network, through its deep structure, automatically learns and extracts deep, abstract features from the input electrical data that reflect differences in underground dielectric electrical properties and the distribution of anomalies. These two feature extraction networks preferably employ network structures tailored to the modal characteristics of their respective data, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or transformers, and their variants, to efficiently extract more concise and discriminative feature representations from high-dimensional raw data.

[0048] After obtaining deep abstract features from seismic and electrical data, the key to this step is to deeply fuse the features of these two modalities using an efficient cross-modal attention mechanism to generate the final fused features. To this end, the previously extracted deep abstract features of seismic data and electrical data are fed together as input to a cross-modal attention module based on a transformer-encoder architecture.

[0049] Within this cross-modal attention module, its core operating mechanism is to project the deep abstract features of one modality into a query matrix Q, and project the deep abstract features of the other modality into a key matrix K and a value matrix V, through a specific, learnable projection transformation. Notably, this mapping relationship can also be reversed, with the query matrix Q generated by the features of the second modality and the key matrix K and value matrix V generated by the features of the first modality, to achieve bidirectional cross-modal information interaction and feature enhancement. This design allows features of one modality to "query" the features of the other modality for the most relevant and complementary parts.

[0050] Using the query matrix Q, key matrix K, and value matrix V generated in this way, the attention-weighted output is calculated by the following formula, which can effectively capture and quantify the intrinsic correlation between the features of the two modalities: Where K T is the transpose of the key matrix K; d k is the dimension of the key vector in the key matrix k, using Scaling is performed to maintain gradient stability; softmax is a normalized exponential function used to convert similarity scores into attention weights; V is the value matrix. The AttentionOutput calculated by this formula is essentially the weighted sum of the value matrix V (i.e., features of one modality) based on its correlation with the query matrix Q (i.e., features of the other modality).

[0051] Based on the weighted attention outputs calculated from one or more applications of the aforementioned attention mechanism, and combined with techniques such as residual connections and layer normalization, a fused feature is ultimately generated. This fused feature does not simply concatenate or average the features of the two modalities. Instead, it intelligently learns the complex dependencies between them through a cross-modal attention mechanism, highlighting complementary information and suppressing redundant or noisy information. Therefore, this fused feature integrates the complementary information of seismic and electrical data, providing a more robust, informative, and discriminative feature foundation for subsequent precise goaf identification and parametric characterization.

[0052] Regarding step S4, in this embodiment, this step aims to reconstruct an accurate underground physical parameter model from abstract features by combining the nonlinear mapping capability of deep learning with geophysical theory and structural priors.

[0053] Specifically, this step first feeds the fused features generated by the cross-modal attention mechanism in step S3 as core input to a pre-configured and optimized deep learning joint inversion network. This deep learning joint inversion network preferably adopts an architecture with strong nonlinear fitting and spatial feature learning capabilities, such as an encoder-decoder (Encoder-Decoder) structure (such as U-Net and its variants), a generative adversarial network (GAN) framework, or a customized neural network structure designed specifically for geophysical inversion problems. These network structures can effectively learn and decode the corresponding subsurface physical parameter distribution from high-dimensional, abstract fused features.

[0054] The core task of the deep learning joint inversion network is to reconstruct the 3D seismic velocity distribution and 3D resistivity distribution of the target area of ​​the open-pit mine in parallel or in an intrinsically coupled manner based on the input fusion features. This means that the network may contain two parallel decoding branches, which are used to output the velocity model and the resistivity model respectively; or a unified decoding structure may be adopted to separate the two different physical parameter models in the deeper layers or output layers of the network. Regardless of the specific structure adopted, the network design aims to use the comprehensive information about the two physical properties contained in the fusion features to infer the spatial distribution of the two models simultaneously or collaboratively. Ultimately, the 3D seismic velocity distribution reconstructed by the deep learning joint inversion network will be output as the final 3D velocity model, and the reconstructed 3D resistivity distribution will be output as the final 3D resistivity model.

[0055] The superior performance and reliability of the results of this deep learning joint inversion network largely rely on the mechanism introduced during its training process, which deeply reflects the idea of ​​physical information guidance. Specifically, the training process of the network is to minimize a designed, multi-objective driven comprehensive loss function L totalAs the core goal. The comprehensive loss function L total It is clearly defined as the weighted sum of multiple key components, and its mathematical expression is as follows: L total =λ data L data +λ phy L phy +λ couple L couple ; Where, L data is the data fitting term. This term ensures that the velocity and resistivity models generated by the network can better fit the original observation data; L phy L is the physical equation constraint term, which directly embeds the known geophysical control equations (such as the wave equation and Poisson equation) into the network optimization, making the generated model consistent at the level of physical laws; couple is the structural coupling constraint term. Given that the structural responses of different physical methods to the same geological body should be spatially correlated, this term utilizes the structural prior constraints across physical properties to invert the results. data ,λ phy ,λ couple They are the non-negative weight coefficients of the data fitting term, the physical equation constraint term, and the structural coupling constraint term, respectively. They are hyperparameters used to balance the relative importance of each constraint in the total loss and need to be optimized through experiments or adaptive strategies.

[0056] The network parameters are iteratively adjusted by optimization algorithms such as gradient descent to minimize L total After sufficient training, the network can learn the complex mapping from fused features to three-dimensional velocity and resistivity models, and the generated model meets the requirements of data fitting, physical law compliance and structural coupling.

[0057] In summary, step S4 realizes the transformation from abstract features to specific three-dimensional geophysical models through the deep learning joint inversion framework and the introduction of a comprehensive loss function guided by physical information, which improves the automation level, efficiency and reliability of the inversion results, and lays a solid foundation for the subsequent identification and characterization of open-pit mine goafs.

[0058] For step S5, in this embodiment, based on the underground three-dimensional velocity model and three-dimensional resistivity model accurately obtained by deep learning joint inversion in the aforementioned step S4, a specially constructed intelligent recognition network is further used to automatically and finely identify and parametrically characterize potential goafs, and finally outputs the recognition results and parametric characterization information that can clearly define the spatial distribution of goafs and quantify their key attributes. This step is a key bridge connecting geophysical imaging and the evaluation of engineering properties of geological bodies. It aims to convert continuous and distributed three-dimensional physical parameter field information into a structured and quantitative description of discrete underground target bodies (i.e., goafs) with clear engineering geological significance, thereby providing direct, accurate and easy-to-understand data support for subsequent engineering practices such as open-pit mine safety production, disaster risk assessment, and resource planning.

[0059] First, the three-dimensional velocity model (storing the seismic wave velocity values ​​of each point in the space in the form of a three-dimensional grid) and the three-dimensional resistivity model (storing the resistivity values ​​of each point in the space in the form of a three-dimensional grid) reconstructed in step S4 are used as key, information-rich input data and are jointly provided to a pre-trained and performance-optimized intelligent recognition network. The network preferably adopts an advanced three-dimensional image segmentation algorithm based on deep learning, such as a convolutional neural network (3D CNN) architecture that is particularly suitable for processing three-dimensional volume data, such as 3DU-Net, V-Net, or other deep network models customized for specific three-dimensional target recognition tasks. These networks, due to their powerful three-dimensional spatial feature learning ability and context understanding ability, can automatically identify and segment three-dimensional spatial regions with geophysical response characteristics of goaf areas (such as specific velocity and resistivity anomalies) from the input multi-modal three-dimensional physical parameter field.

[0060] The intelligent recognition network processes the input three-dimensional model and, through its carefully designed internal multi-layer nonlinear transformation units, performs in-depth, multi-scale feature extraction and cross-modal information fusion analysis on the two three-dimensional data volumes. During the training phase, the network has learned and mastered comprehensive information such as the specific numerical range, typical spatial geometry, boundary characteristics, and physical property differences between the goaf and the surrounding rock, which are commonly presented in the velocity field and resistivity field. It automatically identifies and accurately segments the specific three-dimensional spatial regions of the underground goaf at the voxel level. This process generates a three-dimensional mask or segmentation map that marks the scope of the goaf, clearly defining the spatial position, geometry, and distribution range of each independent goaf, providing a precise spatial basis for subsequent quantitative analysis.

[0061] After each goaf is identified and segmented by the intelligent recognition network, it is parametrically characterized. For each identified goaf g, a set of quantitative descriptors D is calculated and generated. g , which is composed as follows: {ID g,Geometry g ,Physics g}; Where, ID g It is the unique identifier of the goaf, ensuring that each independent goaf object can be accurately tracked and distinguished in subsequent data management and analysis; Geometry g A set of parameters that characterize the geometric characteristics of the goaf, such as the total volume of the goaf, the position of the center of mass in three-dimensional space coordinates, the length and direction of each main axis, the surface area, regularity or irregularity measurement (such as sphericity and density), and the possible average depth of the roof and floor. These parameters together constitute a detailed description of the spatial form, size and distribution of the goaf; Physics g A set of parameters characterizing the average physical properties within the goaf is obtained based on statistical calculations of the corresponding voxel values ​​of the original 3D velocity model and 3D resistivity model within the segmented 3D spatial region of the goaf. These parameters preferably include the internal average seismic velocity and average resistivity values, as well as the standard deviation or coefficient of variation of these physical parameters within the goaf (to reflect the uniformity of its internal physical properties). They may also include the average physical property contrast between the goaf and its adjacent surrounding rock. These physical parameters provide a key quantitative basis for assessing the filling condition, stability, and potential risks of the goaf.

[0062] Finally, the goaf identification result output in step S5 is a digital representation of the three-dimensional spatial position and shape of each goaf (such as a three-dimensional model file, volume data with goaf boundaries marked); the parametric characterization information is a set of quantitative descriptors D provided for each identified goaf, including unique identification, geometric features, and internal average physical properties. g These detailed and accurate goaf identification results and parameterized characterization information provide direct, comprehensive, and highly information-based technical support and decision-making basis for open-pit mine management departments to carry out a series of key tasks, including systematic cataloging of goafs, long-term dynamic stability monitoring, advanced warning and risk assessment of potential geological hazards, scientific formulation of subsidence area management plans, and future mineral resource planning and sustainable mining.

[0063] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for joint detection and identification of open-pit mine goafs by high-density seismic and electrical methods, characterized in that: The following steps are involved: S1. Acquire high-density seismic observation data and high-density electrical observation data for the target area of ​​the open-pit mine; S2. performing data preprocessing on the acquired high-density seismic observation data and high-density electrical observation data to obtain preprocessed seismic data and preprocessed electrical data; S3. performing deep feature extraction on the preprocessed seismic data and the preprocessed electrical data, and fusing the extracted seismic features and electrical features using a cross-modal attention mechanism to obtain fused features; S4. Based on the fusion features, a deep learning joint inversion network is used to perform joint inversion to obtain a three-dimensional velocity model and a three-dimensional resistivity model of the underground. The training process of the joint inversion network adopts a comprehensive loss function including a data fitting term, a physical equation constraint term, and a structural coupling constraint term to guide physical information. S5. Based on the joint inversion, a three-dimensional velocity model and a three-dimensional resistivity model of the underground are obtained, and an intelligent recognition network is used to identify and parameterize the goaf to obtain the recognition result and parameterization information of the goaf.

2. The method for joint detection and identification of open-pit mine goafs by high-density seismic and electrical methods according to claim 1 is characterized in that: The steps of S1 include: Develop high-density seismic data acquisition plans and high-density electrical data acquisition plans based on the geological conditions and exploration tasks of the target area of ​​the open-pit mine. The high-density seismic data acquisition plans and high-density electrical data acquisition plans include observation geometry parameters and data acquisition operation parameters for high-density coverage. According to the developed high-density seismic data acquisition plan, seismic data acquisition operations are carried out in the target area of ​​the open-pit mine to obtain high-density seismic observation data; According to the formulated high-density electrical data acquisition plan, electrical data acquisition operations are carried out in the target area of ​​the open-pit mine to obtain high-density electrical observation data.

3. The method for joint detection and identification of open-pit mine goafs by high-density seismic and electrical methods according to claim 1 is characterized in that: The steps of S2 include: For the acquired high-density seismic observation data, field record editing, observation channel geometry information assignment, noise suppression, static correction and amplitude compensation processing are performed to obtain the pre-processed seismic data; For the acquired high-density electrical observation data, invalid data point elimination, zero point correction, terrain correction and electrode grounding condition check are performed to obtain the pre-processed electrical data.

4. The method for joint high-density seismic and electrical detection and identification of open-pit mine goaf according to claim 1 is characterized in that: In step S3, the step of performing depth feature extraction on the pre-processed seismic data and the pre-processed electrical data includes: Inputting the preprocessed seismic data into a seismic feature extraction deep neural network to extract deep abstract features of the seismic data; The pre-processed electrical method data is input into an electrical method feature extraction deep neural network to extract deep abstract features of the electrical method data.

5. The method for joint high-density seismic and electrical detection and identification of open-pit mine goaf according to claim 1 is characterized in that: In step S3, the step of fusing the extracted seismic features and electrical features using a cross-modal attention mechanism to obtain fused features includes: Inputting the extracted deep abstract features of the seismic data and the deep abstract features of the electrical data into a cross-modal attention module based on a transformer encoder structure; Inside the cross-modal attention module, the query matrix Q is generated by projecting the deep abstract features of one modality, and the key matrix K and value matrix V are generated by projecting the deep abstract features of the other modality, or vice versa, to perform cross-modal interaction calculations; Using the generated query matrix Q, the key matrix K, and the value matrix V, the attention weighted output is calculated using the following formula to capture and quantify the correlation between the two modal features: Where K T is the transpose of the key matrix K; d k is the dimension of the key vector in the key matrix k; softmax is the normalized exponential function; V is the value matrix; Based on the attention weighted output calculated by applying the attention mechanism once or multiple times, a fusion feature is generated, where the fusion feature integrates complementary information of the seismic data and the electrical data.

6. The method for joint high-density seismic and electrical detection and identification of open-pit mine goaf according to claim 1, characterized in that: In step S4, the step of using a deep learning joint inversion network to perform joint inversion based on the fusion features to obtain a three-dimensional velocity model and a three-dimensional resistivity model of the underground includes: Inputting the fused features into a pre-configured deep learning joint inversion network; The deep learning joint inversion network reconstructs the three-dimensional seismic wave velocity distribution and the three-dimensional resistivity distribution of the target area of ​​the open-pit mine in parallel or jointly based on the fusion features; The three-dimensional seismic wave velocity distribution reconstructed by the deep learning joint inversion network is output as a three-dimensional velocity model, and the three-dimensional resistivity distribution is output as a three-dimensional resistivity model.

7. The method for joint high-density seismic and electrical detection and identification of open-pit mine goaf according to claim 1, characterized in that: In step S4, the training goal of the deep learning joint inversion network is to minimize a comprehensive loss function L total , the comprehensive loss function L total Defined as: L total =λ data L data +λ phy L phy +λ couple L couple ; Where, L data is the data fitting term; L phy is the constraint term of the physical equation; L couple is the structural coupling constraint term; data ,λ phy ,λ couple are the non-negative weight coefficients of the data fitting term, the physical equation constraint term and the structural coupling constraint term respectively.

8. The method for joint high-density seismic and electrical detection and identification of open-pit mine goaf according to claim 1, characterized in that: In step S5, the step of obtaining the underground three-dimensional velocity model and three-dimensional resistivity model based on the joint inversion and using an intelligent recognition network to identify and parameterize the goaf includes: providing the three-dimensional velocity model and the three-dimensional resistivity model as input to a pre-trained intelligent recognition network; The intelligent recognition network processes the input to automatically identify and segment the three-dimensional spatial area of ​​the underground goaf; for each goaf identified by the intelligent recognition network, its geometric parameters and internal average physical properties are calculated to form a parametric representation of the goaf.

9. The method for joint detection and identification of open-pit mine goafs by high-density seismic and electrical methods according to claim 1, characterized in that: The identification results and parameterized characterization information of the goaf obtained in step S5 include: The identification result is a digital representation of the three-dimensional spatial position and shape of each goaf; The parameterized characterization information is a set of quantitative descriptors D determined for each identified goaf g. g , the set D g It is composed in the following ways: {ID g ,Geometry g ,Physics g }; Where, ID g It is the unique identifier of the goaf; Geometry g is a set of parameters that characterize the geometric characteristics of the goaf; Physics g It is a set of parameters that characterize the average physical properties inside the goaf.

Images