Intelligent coalfield seismic exploration method based on machine learning

Abstract

The application discloses an intelligent coalfield seismic exploration method based on machine learning, and relates to the technical field of geological exploration.The method processes seismic wave time series data, combines resistivity distribution, constructs multi-source feature expression of residual distribution, imaging energy distribution and resistivity distribution, and realizes spatial recognition of abnormal regions by using machine learning.On the basis of the recognition of the abnormal regions, resistivity gradient distribution, resistivity spatial variation function and imaging energy attenuation rate and other features in each region are further extracted, different abnormal regions such as water-rich structures, mined-out areas and fault fracture zones are classified, after the classification of the types of the abnormal regions, resistivity gradient, resistivity spatial variation function and imaging energy attenuation rate are introduced under a unified velocity model updating framework, the direction, amplitude and spatial change relationship of the velocity updating amount are constrained, and partitioned constraint inversion of the abnormal regions is realized.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, specifically a machine learning-based intelligent coalfield seismic exploration method. Background Technology

[0002] In complex and confined underground environments such as coalfields, taking anomalous structural areas such as water-rich structural zones, goaf areas, and fault fracture zones as examples, these areas often exhibit characteristics such as low velocity, strong attenuation, and enhanced scattering in conventional seismic wave exploration responses.

[0003] On the one hand, due to the strong heterogeneity and strong absorption of coal seams and surrounding rocks, seismic waves are prone to strong scattering, multi-path propagation or rapid energy attenuation during propagation. These complex propagation effects result in amplitude distortion and interface blurring in reflected wave exploration imaging, making it difficult to accurately reflect the spatial structure beneath the coalfield.

[0004] On the other hand, in the underground coalfield environment, different types of anomalous structures exhibit significant similarities in their seismic wave responses. For example, water-rich areas, goaf areas, and fractured zones may all show low velocity or energy attenuation in imaging results, making it difficult to effectively distinguish the types of anomalous structures based solely on seismic data, which can easily lead to uncertainty in the interpretation results. Summary of the Invention

[0005] (a) Technical problems to be solved This invention provides a machine learning-based intelligent coalfield seismic exploration method, which introduces electrical response parameters of coalfield geology and provides physical property constraints for seismic exploration of various types of anomalous geological structures.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based intelligent coalfield seismic exploration method, comprising the following steps: Seismic wave time series data of the coalfield is acquired, reflected wave sequence and transmitted wave travel time information are extracted through wavefield separation, and transient electromagnetic response data is acquired under a unified spatial coordinate reference. Based on the transient electromagnetic response data, resistivity distribution data of the underground medium of the coalfield is inverted. An initial velocity model is constructed using the travel time information, and the reflected wave sequence is migrated and imaged based on the initial velocity model to generate the reflection interface of the coalfield subsurface medium and its imaging energy distribution. Forward modeling is performed based on the initial velocity model to calculate the residual distribution between the simulation data and the seismic wave time series data. The residual distribution, imaging energy distribution, and resistivity distribution are input into a pre-trained anomaly identification model to obtain the spatial anomaly probability distribution. Anomaly regions are extracted by applying connectivity constraints to the high anomaly probability space. Within each of the aforementioned anomalous regions, the resistivity gradient distribution reflecting the region's electrical boundary, the resistivity spatial variability function reflecting heterogeneity, and the imaging energy attenuation rate are calculated. The anomalous regions are then classified into different types, and corresponding constraint inversion strategies are specified. When performing constrained inversion on various types of anomalous regions, the initial velocity at each spatial location is iteratively updated based on the residual distribution within the region. During the velocity update process, the resistivity gradient distribution is introduced to constrain the velocity change trend at each spatial location, and the velocity change continuity within the anomalous region is constrained by the resistivity spatial variability function. At the same time, the velocity update amplitude within the region is adjusted based on the imaging energy decay rate. At the reflection interface location, the spatial change of the interface location before and after the update is calculated, the spatial range of each abnormal area is updated, and the corresponding abnormal structure type is output to obtain the exploration results of the underground abnormal area and geological structure of the coalfield.

[0007] In some feasible embodiments, in the coalfield roadway, based on the target detection range and the geological conditions in the coalfield, a geophone array is deployed under a unified spatial coordinate reference to form a multi-channel seismic observation system covering the target detection range; at the same time, transient electromagnetic transmitting and receiving devices are deployed at corresponding locations. After the system is deployed, seismic waves are generated by the source excitation device according to the preset sampling period. The detector array synchronously collects the seismic wave field data formed by the propagation in the tunnel and synchronously triggers the transient electromagnetic transmitter. The receiver collects the transient electromagnetic response data caused by the change in the conductivity of the underground medium.

[0008] In some feasible embodiments, the seismic wavefield data and transient electromagnetic response data acquired within the sampling period are time-synchronized and spatially registered under a unified spatial coordinate reference. For the seismic wavefield data acquired within the sampling period, after suppressing low-frequency mechanical interference noise and high-frequency pulse noise in the time-frequency domain, wavefield separation is performed on the mixed seismic wavefield to separate the reflected wave, slot wave and transmitted wave within the sampling period.

[0009] In some feasible embodiments, when constructing the initial velocity model, based on the transmitted wave travel time information extracted by wavefield separation, combined with the spatial positional relationship between each detector and the seismic source excitation device, the propagation path of the seismic wave in the underground medium is established, and according to the correspondence between the transmitted wave travel time and the propagation path, the propagation velocity at each spatial location in the underground coalfield is calculated to construct the initial velocity model of the underground coalfield medium. Based on the initial velocity model, the propagation path of the reflected wave sequence extracted by wavefield separation is traced back, and the reflected wave energy received by each detector is mapped to the corresponding spatial location underground in the coalfield. During the mapping process, the reflected wave energy from different detector locations is superimposed to obtain the imaging energy distribution of each underground spatial location. Based on the imaging energy distribution, the spatial location of the reflected energy concentration is identified, and the location of the reflection interface of the underground medium in the coalfield is determined.

[0010] In some feasible embodiments, when training the anomaly recognition model, coalfield area data that has completed geological verification is selected as training samples. Under a unified spatial coordinate reference, the residual distribution features, imaging energy distribution features and resistivity distribution features corresponding to each spatial location are extracted, and the corresponding location is marked as an abnormal geological structure of the coalfield according to the actual geological data. Based on the multi-source features extracted from each spatial location, a training dataset is constructed with the labeled abnormal structure results. The anomaly recognition model is then trained to obtain a mapping relationship that can characterize the multi-source features and the distribution of abnormal structures. This allows the anomaly recognition model to be used as input for multi-source feature data at each spatial location and output the corresponding anomaly probability.

[0011] In some feasible embodiments, under the spatial coordinate reference, based on the anomaly probability distribution of underground spatial coordinate points in the coalfield, spatial coordinate points exceeding a set threshold are filtered, and the filtering results are aggregated based on spatial connectivity to obtain several anomaly regions; the following steps are performed on each anomaly region: Within the abnormal region, based on the resistivity distribution, the resistivity change between each adjacent spatial coordinate point is calculated to obtain the resistivity change rate distribution in space, and the resistivity gradient distribution is characterized by this change rate. Select pairs of spatial coordinate points with different spatial spacing, calculate the resistivity difference between corresponding positions, and statistically analyze the distribution relationship of resistivity difference with spatial spacing to characterize the degree of resistivity variation in space and obtain the resistivity spatial variability function. Based on the imaging energy distribution, the variation of imaging energy with space is calculated along the direction of seismic wave propagation, and the imaging energy attenuation rate is obtained according to the decreasing trend of energy with propagation distance.

[0012] In some feasible embodiments, after calculating and extracting the resistivity gradient distribution, resistivity spatial variability function, and imaging energy attenuation rate of each anomalous region, the anomalous region type classification process is performed, specifically by executing the following classification judgment: When the resistivity in the abnormal region is distributed at medium to low values, the resistivity gradient changes gently, and the imaging energy decay rate is high, the abnormal region is classified as a water-rich abnormal region. When the resistivity change is large, the resistivity spatial variability function value is high, and the imaging energy shows spatial discontinuous attenuation in the abnormal area, the abnormal area is classified as a goaf anomaly. When the resistivity in the anomalous region is distributed at medium to low values ​​and the imaging energy exhibits both local attenuation and local enhancement, the anomalous region is classified as a fault fracture zone anomaly.

[0013] In some feasible embodiments, for areas identified as water-rich anomalies, the initial velocity update amount for each spatial coordinate point is calculated based on the residual distribution within the area; after the update amount is calculated, the gradient direction corresponding to each spatial coordinate point is calculated based on the resistivity gradient distribution extracted within the anomaly area, and the velocity update amount is projected onto this direction to obtain the update component along the resistivity change direction. Based on the imaging energy attenuation rate extracted within the abnormal region, the update component is scaled proportionally to obtain the adjusted velocity update amount; combined with the resistivity spatial variability function extracted within the abnormal region, the velocity update amounts of adjacent spatial coordinate points are subject to difference constraints, so that the update difference between adjacent positions is within a set range; based on the adjusted velocity update amount, the initial velocity model of each spatial coordinate point is updated.

[0014] In some feasible embodiments, for abnormal areas identified as the goaf, the initial velocity update of each spatial coordinate point is calculated based on the residual distribution within the area; Based on the resistivity spatial variation function extracted within the abnormal region, the allowable difference range between corresponding spatial coordinate points within the abnormal region is calculated, and the velocity update amount of each spatial coordinate point is adjusted so that the difference in update amount between different spatial coordinate points meets the corresponding allowable difference range. Based on the imaging energy attenuation rate calculated and extracted within the abnormal region, the velocity update amounts at different spatial coordinate points are proportionally scaled; based on the adjusted velocity update amounts, the initial velocity model at each spatial coordinate point is updated.

[0015] In some feasible embodiments, for the abnormal area determined to be the fault fracture zone, the initial velocity update amount of each spatial coordinate point is calculated based on the residual distribution within the area; Based on the resistivity gradient distribution extracted within the abnormal region, the gradient direction of each spatial coordinate point is calculated, and the velocity update is decomposed into components along the gradient direction and perpendicular to the gradient direction. Combined with the resistivity spatial variability function extracted within the abnormal region, the variation range of each directional component between adjacent spatial coordinate points is restricted. Based on the imaging energy attenuation rate extracted within the abnormal region, the velocity update amount of each spatial coordinate point is proportionally adjusted; the final velocity update amount is synthesized based on the adjusted directional components, and the initial velocity model of each spatial coordinate point is updated accordingly.

[0016] (III) Beneficial Effects: Compared with the prior art, this invention has the following beneficial effects: This invention constructs a multi-source feature system of earthquakes and resistivity, uses machine learning to achieve spatial identification of anomalous areas, and further extracts features such as resistivity gradient distribution, resistivity spatial variation function and imaging energy attenuation rate in each area to classify different anomalous areas, so that areas with similar seismic responses, such as water-rich structures, goaf areas and fault fracture zones, can be distinguished based on electrical and spatial variation characteristics.

[0017] After classifying the abnormal regions, under the unified velocity model update framework, resistivity gradient, resistivity spatial variability function and imaging energy decay rate are introduced for different types of abnormal regions to constrain the direction, amplitude and spatial variation of velocity update, thereby realizing the partitioned constraint inversion of abnormal regions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a machine learning-based intelligent coalfield seismic exploration method provided in an embodiment of the present invention. Figure 2 The following is a schematic diagram illustrating the sequential process of constructing an initial velocity model of coalfield strata based on separated seismic waves and performing forward modeling to calculate the residual distribution in a machine learning-based intelligent coalfield seismic exploration method provided in this embodiment of the invention. Figure 3 This is a flowchart illustrating the process of applying corresponding compensation constraints to various types of selected areas in an intelligent coalfield seismic exploration method based on machine learning, as provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.

[0021] Combination Figures 1 to 3The present invention illustrates a machine learning-based intelligent coalfield seismic exploration method. This method constructs a multi-source feature system of seismicity and resistivity, uses machine learning to achieve spatial identification of anomalous areas, and classifies anomalous areas into types based on resistivity gradient, resistivity spatial variability function and imaging energy attenuation rate. Based on this, it implements zonal constraint inversion and differentiates the velocity update process of different types of anomalous areas.

[0022] The reason for selecting multi-source signals other than seismic waves in the embodiments of the present invention is that, in the coalfield mining environment, single seismic exploration is difficult to reliably distinguish complex structures such as faults, goafs, water-rich areas, and fracture zones. Specifically, seismic wave exploration mainly reflects changes in elastic parameters, while sudden disasters in coal mines often manifest as areas with abnormal conductivity such as water-rich areas and mud and sand fissures, structural voids such as goafs, and discontinuous fracture zones.

[0023] Therefore, transient electromagnetics and resistivity exploration derived from transient electromagnetic inversion are also introduced in the embodiments of this invention to construct a collaborative exploration system based on seismic wave geometric imaging, electromagnetic conductivity anomaly constraint, and resistivity structure compensation. Seismic waves are used to provide the geometric framework for geological exploration. Due to the significant differences in electrical characteristics between water-rich areas and structural regions with marked medium changes, water-rich structures typically correspond to low resistivity distributions, while goaf areas and dry, fractured zones exhibit relatively high resistivity or spatially varied resistivity distributions.

[0024] Therefore, in the embodiments of the present invention, the introduction of transient electromagnetic (TEM) / resistivity information as a constraint can provide a physical property basis independent of seismic response for the identification of the type of anomalous structure, thereby improving the reliability of the identification results.

[0025] After determining the types of data required for seismic exploration, the first step is to execute S10: acquire seismic wave time series data of the coalfield, extract reflected wave sequences and transmitted wave travel time information through wavefield separation, and acquire transient electromagnetic response data under a unified spatial coordinate reference. Based on the transient electromagnetic response data, the resistivity distribution data of the coalfield subsurface medium is inverted. The simultaneous acquisition of these three data points allows for the subsequent inversion of structural and physical (electrical) property information from the same spatial location, enabling the identification of anomaly causes, rather than merely the discovery of anomalies.

[0026] More specifically, in the environment of coalfield roadways, the three types of data can only achieve subsequent joint inversion under the premise of strict spatial alignment. Therefore, before performing the initial data acquisition steps, a unified spatial reference system must first be established. In some embodiments of the present invention, a spatial coordinate framework for the roadway is constructed by performing precise three-dimensional measurements of the roadway, and this framework is used as a reference system for the deployment of the detector array and the transient electromagnetic excitation / response system.

[0027] Under this spatial reference, a multi-channel detector array is deployed along the roadway to form a multi-path wave field sampling network, and transient electromagnetic transmitting and receiving devices are deployed simultaneously at the corresponding locations to construct a conductivity anomaly detection network, while also taking into account the structure of the coal mine roadway.

[0028] After the system deployment is completed, the synchronous acquisition phase begins. First, the detector system, transient electromagnetic system and resistivity acquisition system are synchronized with a unified clock to ensure that the three types of systems work under the same time reference.

[0029] Subsequently, within each preset sampling period, the seismic source excitation device generates seismic wave signals, simultaneously triggering the transient electromagnetic transmission system and response system, enabling each observation and response system to collect underground response information in parallel within the same physical disturbance period.

[0030] During this sampling period, the time series data of the mixed seismic wave field formed by propagation in the tunnel is continuously acquired using the array of detectors deployed above. The transient electromagnetic response curves caused by the underground conductive structure are also acquired synchronously using the transient electromagnetic system and the resistivity acquisition system. Subsequently, the resistivity response distribution data in space is retrieved using the transient electromagnetic system.

[0031] Then, a unified time calibration is performed on the three types of data to ensure that they are strictly aligned in terms of sampling start time, sampling window length and time series, and coordinate mapping is completed under a unified spatial reference, thereby achieving synchronous registration of multi-source data in time and space.

[0032] Finally, the output includes spatially aligned seismic wave time-series data, transient electromagnetic response curves, and resistivity spatial distribution data for each sampling period. In other embodiments of the invention, these can be configured into a unified multi-source input data packet for unified use in subsequent execution steps.

[0033] In summary, the multi-source synchronous acquisition constructed through the above process ensures that each spatial location within the acquisition and exploration area simultaneously possesses a geometric response to the geological structure, a response to the electrical conductivity within the space, and a response to the continuity and integrity of the geological (coal seam medium). These response data are primarily used in subsequent processing for the joint inversion and exploration of the coal seam's structure and properties.

[0034] After the initial data acquisition step (S10), considering that the seismic signals collected in the coal mine roadway are actually a mixed wave field formed by the superposition of multiple waveforms, mainly including the following types of waves: Reflected waves are often used to describe the boundary contour response of geological structures. Channel waves are guided waves used in coal seams, primarily reflecting the continuity of the coal seam. Transmitted waves, because they can propagate across structures, are used to reflect velocity changes within a whole geological structure, thereby distinguishing velocity anomaly zones.

[0035] In addition to the waves mentioned above that are mainly used for geological exploration, in actual underground exploration environments, due to the influence of machine noise, the mixed wave field collected will also be superimposed with mechanical noise caused by fans and transportation equipment. In the time-frequency domain, the noise of these mechanical devices generally manifests as low-frequency and energy-stable continuous noise.

[0036] In addition, there may be superimposed electrical pulse interference from motors, electronic control systems, etc. This type of noise usually manifests as short-duration high-frequency pulses with strong pulse amplitude.

[0037] Therefore, in some embodiments of the present invention, for mechanical low-frequency noise in the mixed wave field, by setting 5Hz-200Hz as the effective seismic signal frequency band, frequency components below 5Hz are removed by high-pass filtering, and wavelet decomposition is performed on the remaining signal to remove low-order approximation coefficients in order to weaken mechanical background vibration.

[0038] For the high-frequency electrical pulse noise doped in, because of its short-duration characteristic, some embodiments use instantaneous amplitude changes exceeding 3 within a time window. The sampling points are replaced, and then median filtering is used to eliminate spikes (high-frequency strong amplitudes) and eliminate transient interference that may be caused by motor starting, etc.

[0039] After eliminating interference noise in the mixed wave field, in order to eliminate amplitude inconsistencies caused by different detector coupling states or differences in the tunnel environment, the signals collected by each system channel are further standardized. Specifically, the wave signals are standardized by energy. This ensures that the waveforms of different sampling devices in different spatial positions and with different sampling periods have a uniform scale, laying the computational foundation for subsequent feature separation and machine learning processing.

[0040] After completing the above denoising and standardization, the next step is to separate the mixed wave field. Here, we first refer to Table 1 below and perform separation processing according to the characteristics of different waves.

[0041] Table 1. Propagation characteristics of each type of wave Based on the table above, in the specific separation stage, the wave field energy distribution is first analyzed in the frequency-wavenumber domain (fk diagram). The slot wave is separated by extracting the linear guided wave. Then, the reflection characteristics of the hyperbola are extracted to separate the reflected wave. Finally, the high-frequency wave is extracted to analyze the transmitted wave.

[0042] For example, in the frequency-wavenumber domain plot (fk plot), clearly observed low-frequency linear bands are identified as slot waves; observed high-frequency linear energy bands are identified as transmitted waves; and the remaining mid-frequency hyperbolic energy bands in the fk plot are identified as emitted waves. Ultimately, datasets for reflected waves, slot waves, and transmitted waves are obtained respectively.

[0043] Regarding transient electromagnetic signals, downhole transient electromagnetic signals are also mainly affected by power grid interference (50Hz and its harmonics) and noise interference from energized equipment. Therefore, during the filtering process, notch filters are set at power frequency and harmonic positions such as 50Hz, 100Hz, and 150Hz in the frequency domain to accurately suppress periodic interference. After filtering, the transient electromagnetic response curve is normalized, including but not limited to normalizing the electromagnetic energy density.

[0044] Since the embodiments of this invention target underground coalfields, after performing the aforementioned wavefield separation, the obtained travel time information of the transmitted waves and the corresponding spatial positions of the seismic source and detector are utilized. Based on this data, the underground space of the coalfield is first divided into regular grids, for example, discretized into several units along the roadway direction and vertical direction, and the propagation path between each pair of seismic sources and detectors is mapped onto these grid units.

[0045] For each propagation path, its total travel time is the sum of the propagation times of each grid cell along the path. That is, the path length and propagation velocity of each cell together determine the propagation time of that segment. By collecting multiple sets of source and detector paths and their corresponding travel time data, a travel time constraint relationship covering the entire region is established, and the velocities of each grid cell are jointly solved to obtain a velocity distribution that satisfies the overall travel time matching relationship, i.e., the initial velocity model.

[0046] After obtaining the initial velocity model, the seismic wave records are spatially located by combining the reflected wave sequence. In some embodiments of the present invention, specifically for each reflected wave signal recorded by the detector, the underground spatial location where the reflection may have occurred is inferred based on its arrival time and the positional relationship between the seismic source and the detector, under the constraints of the velocity model.

[0047] Since the wave propagation path can be simplified to: source (source excitation device), reflection point, detector, the energy of the reflected signal can be mapped to the corresponding spatial point by searching for a location in the underground space that satisfies the propagation time matching relationship.

[0048] Since the same spatial location is often pointed to by multiple paths, the energy from different observation paths is superimposed to form an energy distribution map across the entire spatial grid. The areas of significant energy concentration within this distribution correspond to the locations of reflective interfaces, thus revealing the reflective interface structure of the subsurface medium.

[0049] Regarding the reflection interface obtained from the above inversion, it not only reflects the geometric distribution of the subsurface medium and can be used to correct the structural position in subsequent inversion processes, but the imaging energy distribution reflects the intensity of the seismic wave response in different regions, and its spatial variation characteristics can be further used to calculate energy attenuation. Therefore, in conjunction with the embodiments of this invention, the anomalous geological regions within the coalfield to be identified are understood, as shown in Table 2 below.

[0050] Table 2 Comparison of various geological anomalies and imaging results Before entering the actual identification of various anomaly areas (S30), it is necessary to complete the training of the anomaly identification model, which is built offline based on existing geological verification data.

[0051] First, in coalfield areas where drilling or mining verification has been completed, areas with clear geological results are selected as sample areas. Under a unified spatial coordinate datum, the underground space is divided into grid cells, and multi-source feature data is extracted for each spatial location. These features are the same as those extracted in step S30, specifically calculated using the initial velocity model under a unified spatial coordinate datum.

[0052] The residual distribution obtained from forward modeling reflects the locations where the model and reality are inconsistent; The resistivity distribution obtained from transient electromagnetic inversion reflects the conductivity of the underground medium; Energy distribution obtained from migration imaging.

[0053] Subsequently, based on actual geological data (such as borehole and tunnel detection results), each spatial location is marked, for example, as a normal continuous coal seam / continuous medium area, or a specific abnormal structural area.

[0054] By mapping the spatial location, features, and annotations to a corresponding dataset, a training dataset is constructed and input into a machine learning model for training. This allows the model to establish a mapping relationship between multi-source features and anomaly probabilities. After training, an anomaly detection model is obtained, which can output the corresponding anomaly probability value for feature data at any spatial location.

[0055] In practical identification, if an abnormal region exhibits an overall low resistivity with a gradual change, it primarily indicates that the region is not composed of discrete blocks, but rather a continuous medium. In coalfield environments, the most common continuous low-resistivity medium is water-bearing bodies, because water has significantly higher conductivity than coal seams and surrounding rock.

[0056] Further analysis of spatial variations revealed that if the resistivity did not exhibit significant abrupt changes (small gradient) within the region, it indicates that this low-resistivity state is consistent overall, rather than a result of localized anomalies. Combined with Table 2 above, this is clearly different from the characteristics of strong localized variations found in goaf or fractured zones.

[0057] Furthermore, combining this with seismic imaging, if the reflected energy in the area shows an overall attenuation but still maintains a certain continuity, it indicates that the seismic waves are mainly subjected to absorption during propagation, meaning the reflected energy is gradually dissipated rather than undergoing strong scattering or structural damage. Therefore, in the embodiments of this invention, when the resistivity is distributed at medium to low values, the resistivity gradient changes gently, and the imaging energy attenuation rate is high, the anomalous area is classified as a water-rich anomalous area.

[0058] If, within an anomalous region, the resistivity is not a stable low value but rather varies drastically in space, specifically exhibiting alternating high and low values, and its spatial correlation is confirmed to be very poor (high heterogeneity) through the variogram, this indicates that the region is not a single medium but a mixture of multiple structures.

[0059] In coalfields, the most typical example of this situation is the goaf. Goafs often contain simultaneously: cavities (medium resistivity), collapsed debris (medium resistivity), and localized water accumulation (low resistivity). This situation causes drastic changes in resistivity on a very small scale, thus exhibiting high variability.

[0060] When the resistivity of an abnormal area is generally low to medium, but there are certain fluctuations, and these fluctuations are not as drastic as those in a goaf, but rather have a certain structure, this usually means that the area contains both conductive components (such as fissure water) and retains a certain rock mass structure.

[0061] This characteristic is most typical in fault fracture zones. Fault activity causes cracks and fractures in the rock mass, but does not form complete cavities. Therefore, its electrical properties are characterized by: water-filled cracks (low resistivity) and incompletely collapsed fragments (medium to high resistivity). Overall, it forms a state of medium to low resistivity with structural fluctuations.

[0062] Based on the above criteria for identifying various types of anomalous geological structures, it can be summarized as follows: resistivity is used to determine the physical property type within the region, spatial variation is used to determine the structural state within the region, and finally, imaging energy distribution is used to assist in determining the wavefield response mechanism.

[0063] After determining the type of each anomaly region (S40), the corresponding constraint inversion process is executed. Before explaining the specific constraint inversion, the symbol definitions of the formulas used subsequently will be clearly explained. A unified grid space is set. Perform inversion calculations on the above, and define:

[0064] The initial velocity model to be constructed; This represents the initial velocity update obtained from the seismic residuals. It's important to note that in the inversion constraints for all three types of anomaly regions, the first step is based on the residuals. Calculate the initial velocity update: ; in, This is the step size coefficient; Resistivity distribution; For resistivity gradient (vector); For the unit direction of the resistivity gradient: ; This represents the imaging energy attenuation rate (the larger the value, the stronger the attenuation). It is the spatial variability function of resistivity (characterizing the degree of heterogeneity between two points); For point The set of neighboring points.

[0065] Reference Figure 3 First, let's look at the water-rich area. This type of anomalous area is a continuous water-bearing body. Therefore, in the constrained inversion, it is necessary to update smoothly and avoid lateral jumps.

[0066] Regarding directional constraints, only velocity updates in the direction of electrical change need to be retained: ; Specifically, the update amount is projected onto the direction of resistivity change, so that when the model is updated, the update is only allowed to be adjusted along the direction of the electrical boundary (parallel), thus avoiding lateral noise.

[0067] Regarding the amplitude constraint, scaling is based on energy decay, and the weights need to be defined first: ; ; in, The coefficient set to control the attenuation effect. The larger the value, the higher the corresponding weight. The smaller.

[0068] By utilizing the variogram function to constrain spatial continuity, and focusing on neighborhood points... and ,implement: ; ; in, This is the proportionality coefficient. The smaller, The larger the size, the more continuous and smooth the space.

[0069] For goaf areas, the structures within this type of anomalous area are inconsistent, including but not limited to the presence of cavities, collapses, and water accumulation. Therefore, during constrained inversion, forced smoothing cannot be performed; a permissible difference interval needs to be defined first. In the embodiments of this invention, a variogram function is used to define this interval. ; Used to constrain: ； in, The larger the value, the greater the range of acceptable differences.

[0070] Then, amplitude update control is performed. In some embodiments of the present invention, in order to prevent excessively large updates at once during amplitude updates, point-by-point scaling is generally adopted, i.e., the following is performed: ; final: ; in, This represents a limit on the difference between adjacent values ​​not exceeding [a certain value]. .

[0071] Regarding fault fracture zones, this type of anomalous region generally has a clear fault strike (structural direction). That is, when constraining evolution, it is important to note that changes are possible along the fault direction, but evolution perpendicular to the fault direction must be restricted.

[0072] First, we decompose the direction along the electrical variation (approximately the fault direction): ; In the vertical direction: ; Then anisotropy constraints are applied: ; ; and: ; This can be understood as allowing evolution along the fault direction, but restricting the vertical evolution and diffusion.

[0073] Then energy scaling is performed: ; Finally, the components are recombined.

[0074] It is important to note that the above is only a demonstration of one iteration in the velocity update process. After each iteration, the corresponding seismic response is recalculated and the residual changes are obtained. The direction of subsequent updates is adjusted based on the degree of reduction in the residuals, so that the fitting effect within the anomaly area is gradually improved.

[0075] During the inversion process, the positional changes of the reflective interface are tracked synchronously, and the structural morphology is corrected based on the offset of the interface position before and after the update to reduce imaging errors. After multiple iterations, the updated velocity model and the corresponding reflective interface distribution are obtained.

[0076] During the inversion iteration process, the structural parameters preferentially satisfy the geometric boundary conditions formed by the resistivity constraint, while the physical property parameters converge towards the prior direction of water-rich risk while satisfying the seismic data fitting.

[0077] Through multiple iterations, the model achieves an optimal balance between data consistency and physical property rationality, ultimately outputting an optimized coalfield geological structure model.

[0078] In summary, the entire exploration method involves several steps. In the early stage of identifying anomalous geological structures, seismic wave time series data is processed and combined with resistivity distribution to construct multi-source feature representations of residual distribution, imaging energy distribution, and resistivity distribution. The trained anomaly identification model is then used to assess the probability of anomalies at each spatial location. This transforms the raw geophysical response data into anomalous regions with spatial continuity, realizing the conversion from discrete observation data to spatial structure representation and improving the completeness and stability of anomalous region identification.

[0079] Based on the identification of anomalous areas, the resistivity gradient distribution, resistivity spatial variation function and imaging energy attenuation rate of each area are further extracted to classify different anomalous areas, so that areas with similar seismic responses, such as water-rich structures, goaf areas and fault fracture zones, can be distinguished based on electrical and spatial variation characteristics.

[0080] After classifying the abnormal regions, resistivity gradient, resistivity spatial variability function and imaging energy decay rate are introduced under the unified velocity model update framework for different types of abnormal regions. The direction, amplitude and spatial variation of the velocity update are constrained, so that the inversion results have different update characteristics in each abnormal region, thereby improving the adaptability of the velocity model in abnormal regions.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention should also be included within the scope of protection of the present invention.

Claims

1. A machine learning-based intelligent coalfield seismic exploration method, characterized in that, This includes performing the following steps: Seismic wave time series data of the coalfield is acquired, reflected wave sequence and transmitted wave travel time information are extracted through wavefield separation, and transient electromagnetic response data is acquired under a unified spatial coordinate reference. Based on the transient electromagnetic response data, resistivity distribution data of the underground medium of the coalfield is inverted. An initial velocity model is constructed using the travel time information, and the reflected wave sequence is migrated and imaged based on the initial velocity model to generate the reflection interface of the coalfield subsurface medium and its imaging energy distribution. Forward modeling is performed based on the initial velocity model to calculate the residual distribution between the simulation data and the seismic wave time series data. The residual distribution, imaging energy distribution, and resistivity distribution are input into a pre-trained anomaly identification model to obtain the spatial anomaly probability distribution. Anomaly regions are extracted by applying connectivity constraints to the high anomaly probability space. Within each of the aforementioned anomalous regions, the resistivity gradient distribution reflecting the region's electrical boundary, the resistivity spatial variability function reflecting heterogeneity, and the imaging energy attenuation rate are calculated. The anomalous regions are then classified into different types, and corresponding constraint inversion strategies are specified. When performing constrained inversion on various types of anomalous regions, the initial velocity at each spatial location is iteratively updated based on the residual distribution within the region. During the velocity update process, the resistivity gradient distribution is introduced to constrain the velocity change trend at each spatial location, and the velocity change continuity within the anomalous region is constrained by the resistivity spatial variability function. At the same time, the velocity update amplitude within the region is adjusted based on the imaging energy decay rate. At the reflection interface location, the spatial change of the interface location before and after the update is calculated, the spatial range of each abnormal area is updated, and the corresponding abnormal structure type is output to obtain the exploration results of the underground abnormal area and geological structure of the coalfield.

2. The intelligent coalfield seismic exploration method based on machine learning according to claim 1, characterized in that, Within the coalfield roadways, based on the target detection range and the geological conditions within the coalfield, a geophone array is deployed under a unified spatial coordinate reference to form a multi-channel seismic observation system covering the target detection range; simultaneously, transient electromagnetic transmitting and receiving devices are deployed at corresponding locations. After the system is deployed, seismic waves are generated by the source excitation device according to the preset sampling period. The detector array synchronously collects the seismic wave field data formed by the propagation in the tunnel and synchronously triggers the transient electromagnetic transmitter. The receiver collects the transient electromagnetic response data caused by the change in the conductivity of the underground medium.

3. The intelligent coalfield seismic exploration method based on machine learning according to claim 2, characterized in that, The seismic wavefield data and transient electromagnetic response data collected within the sampling period are time-synchronized and spatially registered under the unified spatial coordinate reference. For the seismic wavefield data acquired within the sampling period, after suppressing low-frequency mechanical interference noise and high-frequency impulse noise in the time-frequency domain, wavefield separation is performed on the mixed seismic wavefield to separate the reflected wave, channel wave and transmitted wave within the sampling period.

4. The intelligent coalfield seismic exploration method based on machine learning according to claim 2, characterized in that, When constructing the initial velocity model, based on the travel time information of the transmitted wave extracted by wave field separation, and combined with the spatial positional relationship between each detector and the source excitation device, the propagation path of the seismic wave in the underground medium is established. Based on the correspondence between the travel time of the transmitted wave and the propagation path, the propagation velocity at each spatial location in the underground coalfield is calculated, and the initial velocity model of the underground medium in the coalfield is constructed. Based on the initial velocity model, the propagation path of the reflected wave sequence extracted by wave field separation is traced back, and the reflected wave energy received by each detector is mapped to the corresponding spatial location underground in the coalfield. During the mapping process, the reflected wave energy from different detector locations is superimposed to obtain the imaging energy distribution of various underground spatial locations. Based on the imaging energy distribution, the spatial location of the concentrated reflected energy is identified, and the location of the reflection interface of the underground medium in the coalfield is determined.

5. The intelligent coalfield seismic exploration method based on machine learning according to claim 1, characterized in that, When training the anomaly recognition model, coalfield area data that has been geologically verified is selected as training samples. Under a unified spatial coordinate reference, the residual distribution features, imaging energy distribution features and resistivity distribution features corresponding to each spatial location are extracted, and the corresponding location is marked as an abnormal geological structure of the coalfield according to the actual geological data. Based on the multi-source features extracted from each spatial location, a training dataset is constructed with the labeled abnormal structure results. The anomaly recognition model is then trained to obtain a mapping relationship that can characterize the multi-source features and the distribution of abnormal structures. This allows the anomaly recognition model to be used as input for multi-source feature data at each spatial location and output the corresponding anomaly probability.

6. The intelligent coalfield seismic exploration method based on machine learning according to claim 5, characterized in that, Under the aforementioned spatial coordinate reference, based on the anomaly probability distribution of underground spatial coordinate points in the coalfield, spatial coordinate points exceeding a set threshold are filtered, and the filtering results are aggregated based on spatial connectivity to obtain several anomaly regions; the following steps are performed on each anomaly region: Within the abnormal region, based on the resistivity distribution, the resistivity change between each adjacent spatial coordinate point is calculated to obtain the resistivity change rate distribution in space, and the resistivity gradient distribution is characterized by this change rate. Select pairs of spatial coordinate points with different spatial spacing, calculate the resistivity difference between corresponding positions, and statistically analyze the distribution relationship of resistivity difference with spatial spacing to characterize the degree of resistivity variation in space and obtain the resistivity spatial variability function. Based on the imaging energy distribution, the variation of imaging energy with space is calculated along the direction of seismic wave propagation, and the imaging energy attenuation rate is obtained according to the decreasing trend of energy with propagation distance.

7. The intelligent coalfield seismic exploration method based on machine learning according to claim 6, characterized in that, After calculating and extracting the resistivity gradient distribution, resistivity spatial variability function, and imaging energy attenuation rate of each anomalous region, the anomalous region type is classified, specifically by performing the following classification judgment: When the resistivity in the abnormal region is distributed at medium to low values, the resistivity gradient changes gently, and the imaging energy decay rate is high, the abnormal region is classified as a water-rich abnormal region. When the resistivity change is large, the resistivity spatial variability function value is high, and the imaging energy shows spatial discontinuous attenuation in the abnormal area, the abnormal area is classified as a goaf anomaly. When the resistivity in the anomalous region is distributed at medium to low values ​​and the imaging energy exhibits both local attenuation and local enhancement, the anomalous region is classified as a fault fracture zone anomaly.

8. The intelligent coalfield seismic exploration method based on machine learning according to claim 7, characterized in that, For areas identified as water-rich anomalies, the initial velocity update of each spatial coordinate point is calculated based on the residual distribution within the area; After the update amount is calculated, based on the resistivity gradient distribution extracted in the abnormal region, the gradient direction corresponding to each spatial coordinate point is calculated, and the velocity update amount is projected in this direction to obtain the update component along the resistivity change direction. Based on the imaging energy attenuation rate extracted within the abnormal region, the update component is scaled proportionally to obtain the adjusted velocity update amount. By combining the resistivity spatial variation function extracted from the abnormal region, the velocity update amount of adjacent spatial coordinate points is subject to a difference limit, so that the update difference between adjacent positions is within a set range. The initial velocity model for each spatial coordinate point is updated based on the adjusted velocity update.

9. The intelligent coalfield seismic exploration method based on machine learning according to claim 7, characterized in that, For abnormal areas identified as the goaf, the initial velocity update of each spatial coordinate point is calculated based on the residual distribution within the area; Based on the resistivity spatial variation function extracted within the abnormal region, the allowable difference range between corresponding spatial coordinate points within the abnormal region is calculated, and the velocity update amount of each spatial coordinate point is adjusted so that the difference in update amount between different spatial coordinate points meets the corresponding allowable difference range. Based on the imaging energy attenuation rate calculated and extracted within the abnormal region, the velocity update amount at different spatial coordinate points is proportionally scaled. The initial velocity model for each spatial coordinate point is updated based on the adjusted velocity update.

10. The intelligent coalfield seismic exploration method based on machine learning according to claim 7, characterized in that, Within the abnormal region identified as the fault fracture zone, the initial velocity update of each spatial coordinate point is calculated based on the residual distribution within the region. Based on the resistivity gradient distribution extracted from the abnormal region, the gradient direction of each spatial coordinate point is calculated, and the velocity update is decomposed into directions to obtain the components along the gradient direction and perpendicular to the gradient direction. By combining the resistivity spatial variation function extracted from the abnormal region, the variation range of each directional component between adjacent spatial coordinate points is restricted; Based on the imaging energy attenuation rate extracted within the abnormal region, the velocity update amount of each spatial coordinate point is proportionally adjusted. The final velocity update is synthesized based on the adjusted directional components, and the initial velocity model of each spatial coordinate point is updated accordingly.

Images