Earthquake intelligent exploration method and system based on multimode fusion

Abstract

The invention relates to the technical field of seismic physical exploration, in particular to an intelligent seismic exploration method and system based on multimode fusion. The method comprises the following steps: collecting multiple types of geophysical observation data, carrying out space registration and time registration, and generating a multimode observation data set; extracting passive source seismological observation data, constructing a resonance response feature tensor, and performing forward calculation by using a resonance peak identification neural network model to generate a resonance feature set; performing feature alignment according to the grid node index to construct a multi-mode fusion feature vector, and performing inversion calculation on the multi-mode fusion feature vector by using a multi-mode fusion inversion model in combination with a learnable modal weight mechanism to obtain a physical property parameter set; and performing segmentation processing and spatial clustering processing on the physical property parameters of each grid node, generating a stratum interface set and an anomalous body set, and obtaining an intelligent seismic exploration result. According to the invention, a seismic exploration mode capable of continuously presenting an underground structure in a complex environment and giving a stable physical property result is realized.

Description

Technical Field

[0001] This invention relates to the field of seismic physical exploration technology, specifically to a seismic intelligent exploration method and system based on multi-mode fusion. Background Technology

[0002] In geophysical exploration, passive source seismic exploration methods are highly feasible in complex surface areas because they do not require external excitation sources. Passive source seismic records contain the response of the subsurface medium to continuous natural vibrations, and the resonance phenomenon can reflect the wave impedance structure of the subsurface medium. Traditional resonance methods usually identify resonance peaks through manual or semi-automatic analysis of spectral energy to infer the wave impedance distribution of the subsurface medium. At the same time, other geophysical methods such as microseismic, shallow seismic, and gravity observations are often used to provide lateral structure or density information to enhance the reliability of exploration interpretation. In actual exploration using passive source seismic methods, there are still problems such as the resonance peak of the same stratum appearing in different locations with varying strengths, frequency drift, or even complete disappearance due to inconsistent natural vibration composition received at different measuring points. In complex surface environments, the vibration background differences caused by nearby roads, construction machinery, or local topographical differences can further amplify the aforementioned problems of the resonance peak of the same stratum appearing in different locations with varying strengths, frequency drift, or even complete disappearance. Summary of the Invention

[0003] The purpose of this invention is to provide a seismic intelligent exploration method and system based on multi-mode fusion, so as to solve the problem mentioned in the background art that the resonance peak of the same stratum exhibits different strengths, frequency drifts, or even disappears completely at different locations.

[0004] To achieve the above objectives, the technical solution of the present invention is: a seismic intelligent exploration method based on multi-mode fusion, comprising: S1. Collect multiple types of geophysical observation data within the target exploration area, construct a three-dimensional exploration grid based on preset horizontal and depth coordinates, and perform spatial and temporal registration of the multiple types of geophysical observation data with the grid nodes of the three-dimensional exploration grid according to the observation location to generate a multi-mode observation dataset. S2. Extract passive source seismic observation data corresponding to each grid node from the multi-mode observation dataset, perform frequency domain transformation on the passive source seismic observation data, and perform frequency resampling according to the preset frequency sampling interval. Calculate the spectral energy distribution of each grid node at each frequency sampling point, construct the resonance response feature tensor, and use the resonance peak recognition neural network model to perform forward calculation to generate a set of resonance features. S3. Align the fields of the resonance feature set and other observation data in the multi-mode observation dataset, excluding passive source seismic observation data, according to the grid node index to construct a multi-mode fusion feature vector. Use the multi-mode fusion inversion model and combine the learnable mode weight mechanism to perform inversion calculation on the multi-mode fusion feature vector to obtain the set of physical property parameters. S4. Based on the set of physical property parameters, perform segmentation and spatial clustering processing on the physical property parameters of each grid node to generate a set of stratigraphic interfaces and anomaly bodies, and obtain the results of intelligent seismic exploration.

[0005] Preferably, in step S2, the resonant response feature tensor is a multi-dimensional array structure consisting of grid node index as the first dimension, frequency sampling point index as the second dimension, and each spectral energy statistical field as the third dimension. It is used to uniformly and structurally express the spectral energy distribution of each grid node at each frequency sampling point. The construction method of the resonant response feature tensor is as follows: after frequency domain transformation of the passive source seismic observation data corresponding to each grid node, resampling is performed according to a preset frequency sampling interval to obtain the spectral energy field of each frequency sampling point. The spectral energy field is arranged according to the grid node index and the frequency sampling point index, and combined in a fixed field order to form the resonant response feature tensor.

[0006] Preferably, in S2, the resonant peak recognition neural network model is based on a neural network structure that takes tensor slices of the resonant response feature tensor as input and outputs the resonant peak frequency field and the resonant peak confidence field. It is used to perform peak recognition and confidence estimation on the resonant response features corresponding to each grid node. The structure of the resonant peak recognition neural network model is as follows: a multi-layer feature extraction network with frequency as the main feature extraction dimension is used as the input layer to perform frequency domain feature encoding on the input tensor slices; a mapping network with fully connected layers as the output layer is used to map the encoded features to the resonant peak frequency field and the resonant peak confidence field. The specific method for the resonant peak recognition neural network model to perform forward calculation on the resonant response feature tensor is as follows: the resonant response feature tensor is split into several tensor slices according to the grid node index; each tensor slice is input into the neural network model to perform forward calculation to obtain the corresponding resonant peak frequency field and the resonant peak confidence field; and the calculation results of each grid node are combined according to the index to generate a resonant feature set.

[0007] Preferably, in step S2, the resonance feature set includes a resonance peak frequency field and a resonance peak confidence field corresponding to each grid node; wherein, the resonance peak frequency field is a resonance peak frequency value output from the resonance peak recognition neural network model, used to characterize the main resonance frequency corresponding to each grid node; the resonance peak confidence field is a confidence value output from the resonance peak recognition neural network model, used to characterize the stability of the corresponding resonance peak frequency; the resonance peak confidence field is obtained based on a confidence estimation mechanism; wherein, the confidence estimation mechanism includes: statistically analyzing the energy change trend of the resonance response feature tensor at continuous frequency sampling points, combining and calculating the energy gradient field, frequency local consistency field, and outlier suppression field near the resonance peak frequency, and inputting the above fields into the resonance peak recognition neural network model according to preset rules to form a confidence output field.

[0008] Preferably, in step S3, the multi-mode fusion feature vector is a one-dimensional vector formed by concatenating the resonance feature fields corresponding to each grid node with other observation data fields in the multi-mode observation dataset in a fixed order. It is used to uniformly represent the features of multiple types of observation information at the same grid node. The data structure of the multi-mode fusion feature vector is a set of vectors arranged by grid node index. Each vector is composed of the resonance peak frequency field, the resonance peak confidence field, and other observation data fields in sequence. The construction method of the multi-mode fusion feature vector is as follows: the various observation data fields in the multi-mode observation dataset are aligned according to the grid node index, and the aligned fields are arranged and combined in a fixed field order to form the multi-mode fusion feature vector.

[0009] Preferably, in S3, the learnable modal weighting mechanism is a set of parameters for setting corresponding modal weight parameters for various observation features in the multi-mode fusion feature vector and updating each modal weight parameter during the inversion process, used to adjust the contribution of different observation features in the inversion calculation; the learnable modal weighting mechanism is specifically as follows: in the multi-mode fusion inversion model, a first modal weight parameter is set for the resonance feature field, a second modal weight parameter and subsequent modal weight parameters are set for other observation feature fields, each modal weight parameter is combined with the feature value of the corresponding field according to a preset operation rule, and each modal weight parameter is updated based on the inversion error during the inversion model training process.

[0010] Preferably, in S3, the multi-mode fusion inversion model is an inversion calculation model constructed based on multi-mode fusion feature vectors and a learnable mode weighting mechanism. It is used to perform inversion calculations on the multi-mode fusion feature vectors of each grid node to obtain the wave impedance parameter field and density parameter field corresponding to the grid node. The structure of the multi-mode fusion inversion model is defined as follows: an input layer receives the multi-mode fusion feature vectors, a feature weighting layer with embedded mode weighting parameters performs mode weighting processing on the input vectors, a hidden layer performs feature mapping operations and obtains the relevant feature expressions of physical property parameters, and an output layer generates the wave impedance parameter field and density parameter field. The multi-mode fusion inversion model performs inversion calculations on the multi-mode fusion feature vectors in conjunction with the learnable mode weighting mechanism, specifically as follows: the multi-mode fusion feature vectors are input into the feature weighting layer of the inversion model, each feature field is weighted according to the mode weighting parameters, and then input into the hidden layer to perform inversion mapping, and the wave impedance parameter field and density parameter field are generated in the output layer.

[0011] Preferably, in step S3, the set of physical property parameters is a data set with grid node index as the key and corresponding wave impedance parameter field and density parameter field as the value, including the wave impedance parameter field of each grid node and the density parameter field of each grid node.

[0012] On the other hand, the present invention provides a seismic intelligent exploration system based on multi-mode fusion, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the aforementioned seismic intelligent exploration method based on multi-mode fusion.

[0013] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: 1. In this invention, a resonance response feature tensor is constructed based on passive source seismic observation data and a neural network model is used to identify resonance peaks. This can clearly obtain the resonance peak positions of each grid node even when the vibration background of different measuring points is inconsistent. This avoids the phenomenon of missing, shifted or discontinuous resonance peaks caused by differences in background noise, and enables the resonance information of the same stratum to be continuously expressed throughout the entire measuring area. 2. In this invention, by unifying the resonance characteristics with other geophysical observation data into a multi-mode fusion feature vector, and introducing a learnable mode weighting mechanism into the inversion model, the reliability differences of different observation data are automatically adjusted, so that the wave impedance parameters and density parameters obtained by inversion remain stable in the lateral variation region and the low signal-to-noise ratio region, thereby obtaining physical property distribution results that are more consistent with the real geological structure. Attached Figure Description

[0014] Figure 1 This is a flowchart of one embodiment of the present invention. Detailed Implementation

[0015] Example 1, as Figure 1 As shown, the specific implementation steps of the intelligent seismic exploration method based on multi-mode fusion proposed in this invention are as follows: S1. Collect multiple types of geophysical observation data within the target exploration area, construct a three-dimensional exploration grid based on preset horizontal and depth coordinates, and perform spatial and temporal registration of the multiple types of geophysical observation data with the grid nodes of the three-dimensional exploration grid according to the observation location to generate a multi-mode observation dataset. S2. Extract passive source seismic observation data corresponding to each grid node from the multi-mode observation dataset, perform frequency domain transformation on the passive source seismic observation data, and perform frequency resampling according to the preset frequency sampling interval. Calculate the spectral energy distribution of each grid node at each frequency sampling point, construct the resonance response feature tensor, and use the resonance peak recognition neural network model to perform forward calculation to generate a set of resonance features. S3. Align the fields of the resonance feature set and other observation data in the multi-mode observation dataset, excluding passive source seismic observation data, according to the grid node index to construct a multi-mode fusion feature vector. Use the multi-mode fusion inversion model and combine the learnable mode weight mechanism to perform inversion calculation on the multi-mode fusion feature vector to obtain the set of physical property parameters. S4. Based on the set of physical property parameters, perform segmentation and spatial clustering processing on the physical property parameters of each grid node to generate a set of stratigraphic interfaces and anomaly bodies, and obtain the results of intelligent seismic exploration.

[0016] In this embodiment S1, the target exploration area refers to the spatial range within which information on the subsurface medium structure needs to be obtained. This spatial range is usually determined based on the alignment or site layout of the actual exploration task. Multiple types of geophysical observation data refer to the original observation records obtained within the target exploration area through different observation methods, including passive source seismic observation data, shallow micromotion observation data, surface waveform record data, and other observation data fields used to supplement the lateral changes of the subsurface structure. The preset horizontal coordinates and depth coordinates are coordinate sequences used to establish a spatial grid, where the horizontal coordinates are used to discretize the planar range of the exploration area, and the depth coordinates are used to discretize the vertical range.

[0017] In this embodiment S1, the three-dimensional exploration grid is a three-dimensional grid structure formed on the basis of the above coordinate sequence. Each grid node is determined by the corresponding horizontal coordinate and depth coordinate. The three-dimensional exploration grid can be constructed based on the preset horizontal coordinate and depth coordinate by adopting the regular grid division method or the irregular grid division method commonly used in the existing seismic exploration field. By setting fixed step sizes in the horizontal and vertical directions to regularly divide the spatial domain, a set of grid nodes with fixed index relationships is formed. When spatially registering multiple types of geophysical observation data with the grid nodes of the three-dimensional exploration grid according to the observation location, the coordinates of the observation points recorded in the observation data are matched with the location of the nearest grid node, so that data from different sources belong to a unified spatial index. Time registration refers to aligning the records with the time axis according to the sampling time or the start and end time of the acquisition of various types of observation data, so that different observation records of the same grid node can be processed together under the same time reference. The multi-mode observation dataset refers to the data set formed by the structured storage of all registered observation data based on the three-dimensional exploration grid as the index, including the passive source seismic observation data field, shallow micromotion observation data field, surface waveform field and other observation data fields corresponding to each grid node.

[0018] In this embodiment S2, the resonant response feature tensor is a multi-dimensional array structure consisting of grid node index as the first dimension, frequency sampling point index as the second dimension, and each spectral energy statistical field as the third dimension. It is used to uniformly and structurally express the spectral energy distribution of each grid node at each frequency sampling point. The construction method of the resonant response feature tensor is as follows: after frequency domain transformation of the passive source seismic observation data corresponding to each grid node, resampling is performed according to a preset frequency sampling interval to obtain the spectral energy field of each frequency sampling point. The spectral energy field is arranged according to the grid node index and the frequency sampling point index, and combined in a fixed field order to form the resonant response feature tensor.

[0019] In this embodiment S2, extracting passive source seismic observation data corresponding to each grid node from the multi-mode observation dataset refers to reading the original seismic records that match the spatial location of the node based on the grid node index in the multi-mode observation dataset. The passive source seismic observation data are continuous seismic records generated by natural environmental vibrations without the use of artificial excitation sources. They are used to reflect the response characteristics of the subsurface medium under natural vibrations, including three-component seismic waveform data recorded in time series form. The preset frequency sampling interval is a parameter that is fixedly set for the frequency step size between frequency domain sampling points. It is usually determined based on the recording sampling rate and the target exploration depth range, so that the frequency domain energy forms a sampling point sequence at fixed intervals within the specified frequency range.

[0020] In this embodiment S2, the method for calculating the spectral energy distribution of each grid node at each frequency sampling point is as follows: A frequency domain transformation is performed on the passive source seismic observation data of the corresponding node, converting the time-domain waveform into a frequency-domain representation. Then, values ​​are taken at a preset frequency sampling interval, and the energy value corresponding to each frequency sampling point is calculated. The spectral energy distribution is a set of energy values ​​recorded with frequency sampling points as the index, used to describe the energy changes of each node at different frequencies. The data structure is a one-dimensional energy sequence arranged in order of frequency sampling points. The frequency domain transformation of the passive source seismic observation data corresponding to each grid node can be performed using existing fast Fourier transform technology to map the time-series signal to the frequency domain. Resampling at the preset frequency sampling interval uses a fixed-interval sampling method, so that the spectrum forms a unified sampling point sequence at a given interval. Obtaining the spectral energy field of each frequency sampling point refers to reading the energy value corresponding to each sampling point in the frequency domain representation and forming a field set. The spectral energy field is an energy value field indexed by the sampling frequency, used to construct the energy dimension of the subsequent tensor structure.

[0021] In this embodiment S2, the method for arranging the spectral energy field according to the grid node index and the frequency sampling point index is as follows: first, determine the first dimension according to the grid node index, then determine the second dimension according to the frequency sampling point index, and store the energy field in the corresponding position according to the fixed field order, wherein the fixed field order is the energy field arranged in order of frequency sampling points from low to high; the method for combining the resonant response feature tensor under the fixed field order is as follows: with the grid node as the first dimension, the frequency sampling point as the second dimension, and the energy field order as the third dimension, combine the energy sequences of all nodes into a unified three-dimensional array structure according to the above rules, and then input the three-dimensional array into the resonant peak recognition neural network model to perform forward calculation to generate a resonant feature set.

[0022] In this embodiment S2, the resonant peak recognition neural network model is based on a neural network structure that takes tensor slices of the resonant response feature tensor as input and outputs the resonant peak frequency field and the resonant peak confidence field. It is used to perform peak recognition and confidence estimation on the resonant response features corresponding to each grid node. The structure of the resonant peak recognition neural network model is as follows: a multi-layer feature extraction network with frequency dimension as the main feature extraction dimension is used as the input layer to perform frequency domain feature encoding on the input tensor slices; a mapping network with fully connected layers as the output layer is used to map the encoded features to the resonant peak frequency field and the resonant peak confidence field. The specific method of the resonant peak recognition neural network model for forward calculation of the resonant response feature tensor is as follows: the encoded features are mapped to the resonant peak frequency field and the resonant peak confidence field; each tensor slice is input into the neural network model to perform forward calculation to obtain the corresponding resonant peak frequency field and the resonant peak confidence field; and the calculation results of each grid node are combined by index to generate a resonant feature set.

[0023] In this embodiment S2, the tensor slice of the resonant response feature tensor refers to a two-dimensional subarray formed by obtaining the energy sequence of a node at all frequency sampling points in the constructed three-dimensional array structure with a single grid node as the first-dimensional index and fixing the index. This subarray is used as the input unit of the neural network model. The method for performing peak identification and confidence estimation on the resonant response features corresponding to each grid node is as follows: each tensor slice is input into the resonant peak identification neural network model, so that the model infers the frequency position of the resonant peak based on the energy change relationship corresponding to different frequency sampling points and generates a confidence value for the stability of the frequency position. The frequency dimension refers to the dimension arranged according to the frequency sampling points in the tensor structure. This dimension reflects the energy value sequence at each frequency. The reason for using a multi-layer feature extraction network with the frequency dimension as the main feature extraction dimension as the input layer is that the frequency dimension can directly characterize the energy change pattern of the resonant peak formation. Using the frequency sequence as a continuous feature input helps the model extract the local peak structure and the neighborhood energy relationship.

[0024] In this embodiment S2, the method for performing frequency domain feature encoding on the input tensor slice is as follows: local features are extracted layer by layer in the frequency dimension through a multi-layer convolutional network or other structures that can be used to extract one-dimensional sequence features, and encoded feature representations are formed, so that the encoded features can represent the energy differences and local change patterns at the frequency. The reason for using a mapping network with fully connected layers as the output layer is that the fully connected structure can map the encoded sequence features into fixed-dimensional output fields, which are used to form the resonant peak frequency field and the resonant peak confidence field. The method for mapping the encoded features into the resonant peak frequency field is as follows: an output node for regressing frequency values ​​is set in the output layer, so that the mapping result is a frequency value representing the peak position. The method for mapping the encoded features into the resonant peak confidence field is as follows: a node for outputting confidence values ​​is set in the output layer, so that the mapping result is a confidence value representing the stability of the peak.

[0025] In this embodiment S2, the specific method for inputting each tensor slice into the neural network model to perform forward calculation to obtain the corresponding resonance peak frequency field and resonance peak confidence field is as follows: read the tensor slice of each grid node in sequence and input it into the neural network model to perform forward calculation once to obtain the corresponding resonance peak frequency value and confidence value, and then store the above two fields into the corresponding positions according to the grid node index to form a resonance feature set.

[0026] In this embodiment S2, the resonance feature set includes a resonance peak frequency field and a resonance peak confidence field corresponding to each grid node. The resonance peak frequency field is a resonance peak frequency value output from the resonance peak recognition neural network model, used to characterize the main resonance frequency corresponding to each grid node. The resonance peak confidence field is a confidence value output from the resonance peak recognition neural network model, used to characterize the stability of the corresponding resonance peak frequency. The resonance peak confidence field is obtained based on a confidence estimation mechanism. This confidence estimation mechanism includes: statistically analyzing the energy change trend of the resonance response feature tensor at continuous frequency sampling points; combining and calculating the energy gradient field, frequency local consistency field, and outlier suppression field near the resonance peak frequency; and inputting these fields into the resonance peak recognition neural network model according to preset rules to form a confidence output field.

[0027] In this embodiment S2, the main resonant frequency corresponding to each grid node refers to the position in the resonant response characteristics of that node where the energy change is most obvious and conforms to the resonant characteristics. Its value is given by the resonant peak frequency field output by the resonant peak recognition neural network model after performing forward calculation on the input tensor slice. In the credibility estimation mechanism, the method for statistically analyzing the energy change trend of the resonant response feature tensor at continuous frequency sampling points is as follows: read the energy value in the tensor slice in the order of frequency sampling points and calculate the energy difference between adjacent sampling points to form an energy trend sequence. Then, based on the trend sequence, determine whether the energy shows a continuous transition relationship that conforms to the resonant characteristics. The method for combining the energy gradient field, frequency local consistency field, and outlier suppression field near the resonant peak frequency is as follows: select a fixed range of frequency neighborhood near the frequency position corresponding to the main resonant frequency and calculate the energy gradient field to represent the local energy change rate. Then, calculate the frequency local consistency field to reflect the consistency of the energy pattern in the neighborhood. At the same time, extract the outlier suppression field based on the energy trend sequence to weaken the abrupt energy points that do not conform to the resonant characteristics. Then, input the above three fields into the credibility output node of the resonant peak recognition neural network model according to the preset rules to form a credibility value.

[0028] In this embodiment S3, the multi-mode fusion feature vector is a one-dimensional vector formed by concatenating the resonance feature fields corresponding to each grid node with other observation data fields in the multi-mode observation dataset in a fixed order. It is used to uniformly represent the features of multiple types of observation information at the same grid node. The data structure of the multi-mode fusion feature vector is a set of vectors arranged by grid node index. Each vector is composed of the resonance peak frequency field, the resonance peak confidence field, and other observation data fields in sequence. The construction method of the multi-mode fusion feature vector is as follows: the various observation data fields in the multi-mode observation dataset are aligned according to the grid node index, and the aligned fields are arranged and combined in a fixed field order to form the multi-mode fusion feature vector.

[0029] In this embodiment S3, the multi-mode fusion feature vector is a one-dimensional vector formed by concatenating the resonance feature fields corresponding to each grid node with other observation data fields in the multi-mode observation dataset in a fixed order. The fixed order refers to the order in which the resonance peak frequency field is the first field, the resonance peak confidence field is the second field, and other observation data fields are arranged in order according to their field numbers recorded in the multi-mode observation dataset. The method for feature alignment processing of various observation data fields in the multi-mode observation dataset according to the grid node index is as follows: in the three-dimensional exploration grid, the same grid node index is used as a unified index basis. The corresponding field values ​​of the resonance peak frequency field, resonance peak confidence field, and other observation data fields corresponding to the node are read according to the index, and the reading results form a field set with the grid node index as the key. The method for arranging and combining the aligned fields in a fixed field order to form the multi-mode fusion feature vector is as follows: the field values ​​are extracted in a fixed order from the field set corresponding to each grid node, and the extracted field values ​​are stored in a one-dimensional array structure in order, so that the index order of the array structure corresponds one-to-one with the field order, thereby forming a multi-mode fusion feature vector for subsequent inversion calculation.

[0030] In this embodiment S3, the learnable modal weighting mechanism is a set of parameters that sets corresponding modal weight parameters for various observation features in the multi-mode fusion feature vector and updates each modal weight parameter during the inversion process. It is used to adjust the contribution of different observation features in the inversion calculation. The learnable modal weighting mechanism is as follows: in the multi-mode fusion inversion model, a first modal weight parameter is set for the resonance feature field, and a second modal weight parameter and subsequent modal weight parameters are set for other observation feature fields. Each modal weight parameter is combined with the feature value of the corresponding field according to a preset operation rule, and each modal weight parameter is updated based on the inversion error during the inversion model training process.

[0031] In this embodiment S3, setting corresponding modal weight parameters for various observation features in the multi-mode fusion feature vector refers to assigning one or a set of weight coefficients corresponding one-to-one with each field to the resonance peak frequency field, resonance peak confidence field, and other observation data fields in the multi-mode observation dataset in the multi-mode fusion inversion model. This is used to weight the values ​​of each field before inversion calculation. Modal weight parameters are one of the model parameters participating in the inversion calculation, including first-type modal weight parameters applied to the resonance feature field and second-type and subsequent modal weight parameters applied to various other observation feature fields. All modal weight parameters are stored in the parameter set of the multi-mode fusion inversion model in scalar or vector form. Updating each modal weight parameter during the inversion process refers to... After performing inversion calculations on the multi-mode fusion feature vectors and outputting the physical property parameters, the multi-mode fusion inversion model compares the output physical property parameters with the target physical property parameter data used for training. Based on the error calculation results between the two, the model parameter set, including the modal weight parameters, is adjusted. The learnable modal weight mechanism refers to using the modal weight parameters as trainable parameters within the multi-mode fusion inversion model. In each inversion training iteration, the modal weight parameters are updated through the error backpropagation algorithm or other existing parameter update algorithms. This allows the resonant feature field and various other observation feature fields to form a weight allocation that matches their contribution to the physical property parameters after multiple training iterations. Thus, the weighted processing of the multi-mode fusion feature vectors and the joint inversion of physical property parameters are completed within a unified inversion framework.

[0032] In this embodiment S3, the multi-mode fusion inversion model is an inversion calculation model constructed based on multi-mode fusion feature vectors and a learnable mode weighting mechanism. It is used to perform inversion calculations on the multi-mode fusion feature vectors of each grid node to obtain the wave impedance parameter field and density parameter field corresponding to the grid node. The structure of the multi-mode fusion inversion model is defined as follows: an input layer receives the multi-mode fusion feature vectors, a feature weighting layer with embedded mode weighting parameters performs mode weighting processing on the input vectors, a hidden layer performs feature mapping operations and obtains the relevant feature expressions of physical property parameters, and an output layer generates the wave impedance parameter field and density parameter field. The multi-mode fusion inversion model performs inversion calculations on the multi-mode fusion feature vectors in conjunction with the learnable mode weighting mechanism, specifically as follows: the multi-mode fusion feature vectors are input into the feature weighting layer of the inversion model, each feature field is weighted according to the mode weighting parameters and then input into the hidden layer for inversion mapping, and the wave impedance parameter field and density parameter field are generated in the output layer.

[0033] In this embodiment S3, the method for performing inversion calculations on the multi-mode fusion feature vectors of each grid node to obtain the wave impedance parameter field and density parameter field corresponding to the grid node is as follows: the multi-mode fusion feature vectors are input into the multi-mode fusion inversion model, and the model performs weighted processing, feature mapping operations, and output generation on the input features; wherein, the wave impedance parameter field is a parameter sequence that records the magnitude of the impedance of the underground medium to the incident wave propagation with the grid node as an index, and is used to characterize the medium velocity and density correlation characteristics of each grid node; the density parameter field is a parameter sequence that records the density values ​​of the underground medium with the grid node as an index, and is used to characterize the physical property changes of each underground node; the feature weighting layer with embedded modal weight parameters is a weighting operation structure located after the input layer in the multi-mode fusion inversion model, which is used to combine the learnable modal weight parameters with the corresponding feature fields in the input vector field by field.

[0034] In this embodiment S3, the method of using a feature weighting layer with embedded modal weight parameters to perform modal weighting processing on the input vector is as follows: the values ​​of each feature field in the input vector are read in a fixed field order, and the values ​​are multiplied with the corresponding modal weight parameters or other existing combination operations that can be used for weighting processing. The result of the combination operation forms a weighted feature vector and is passed to the hidden layer of the model. The method of using the hidden layer to perform feature mapping operation and obtain the feature expression related to physical property parameters is as follows: after the weighted feature vector is input into the hidden layer, a series of feature transformations are performed on the input vector through the nonlinear feature mapping unit set inside the hidden layer, so that the input features form an intermediate expression that can reflect the value characteristics of the wave impedance parameter field and the density parameter field in multiple mapping steps. Then, the intermediate expression is input into the output layer to generate the wave impedance parameter field and density parameter field corresponding to the grid node.

[0035] In this embodiment S3, the physical property parameter set is a data set with grid node index as key and corresponding wave impedance parameter field and density parameter field as value, including the wave impedance parameter field of each grid node and the density parameter field of each grid node.

[0036] Example 2: The present invention proposes a multi-mode fusion-based intelligent seismic exploration system, which is applied to the multi-mode fusion-based intelligent seismic exploration method proposed in Example 1. It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the multi-mode fusion-based intelligent seismic exploration method in Example 1.

[0037] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A seismic intelligent exploration method based on multi-mode fusion, characterized in that, Includes the following steps: S1. Collect multiple types of geophysical observation data within the target exploration area, construct a three-dimensional exploration grid based on preset horizontal and depth coordinates, and perform spatial and temporal registration of the multiple types of geophysical observation data with the grid nodes of the three-dimensional exploration grid according to the observation location to generate a multi-mode observation dataset. S2. Extract passive source seismic observation data corresponding to each grid node from the multi-mode observation dataset, perform frequency domain transformation on the passive source seismic observation data, and perform frequency resampling according to the preset frequency sampling interval. Calculate the spectral energy distribution of each grid node at each frequency sampling point, construct the resonance response feature tensor, and use the resonance peak recognition neural network model to perform forward calculation to generate a set of resonance features. S3. Align the fields of the resonance feature set and other observation data in the multi-mode observation dataset, excluding passive source seismic observation data, according to the grid node index to construct a multi-mode fusion feature vector. Use the multi-mode fusion inversion model and combine the learnable mode weight mechanism to perform inversion calculation on the multi-mode fusion feature vector to obtain the set of physical property parameters. S4. Based on the set of physical property parameters, perform segmentation and spatial clustering processing on the physical property parameters of each grid node to generate a set of stratigraphic interfaces and anomaly bodies, and obtain the results of intelligent seismic exploration.

2. The intelligent seismic exploration method based on multi-mode fusion according to claim 1, characterized in that: In S2, the resonant response feature tensor is a multi-dimensional array structure consisting of grid node index as the first dimension, frequency sampling point index as the second dimension, and each spectral energy statistical field as the third dimension. It is used to uniformly and structurally express the spectral energy distribution of each grid node at each frequency sampling point. The construction method of the resonant response feature tensor is as follows: after frequency domain transformation of the passive source seismic observation data corresponding to each grid node, resampling is performed according to a preset frequency sampling interval to obtain the spectral energy field of each frequency sampling point. The spectral energy field is arranged according to the grid node index and the frequency sampling point index, and combined in a fixed field order to form the resonant response feature tensor.

3. The intelligent seismic exploration method based on multi-mode fusion according to claim 2, characterized in that: In S2, the resonant peak recognition neural network model is a neural network structure based on tensor slices of the resonant response feature tensor as input and resonant peak frequency field and resonant peak confidence field as output. It is used to perform peak recognition and confidence estimation on the resonant response features corresponding to each grid node. The structure of the resonant peak recognition neural network model is as follows: a multi-layer feature extraction network with frequency dimension as the main feature extraction dimension is used as the input layer to perform frequency domain feature encoding on the input tensor slices; a mapping network with fully connected layers as the output layer is used to map the encoded features to the resonant peak frequency field and resonant peak confidence field. The specific method of the resonant peak recognition neural network model for forward calculation of the resonant response feature tensor is as follows: the resonant response feature tensor is split into several tensor slices according to the grid node index, each tensor slice is input into the neural network model to perform forward calculation to obtain the corresponding resonant peak frequency field and resonant peak confidence field, and the calculation results of each grid node are combined according to the index to generate a resonant feature set.

4. The intelligent seismic exploration method based on multi-mode fusion according to claim 3, characterized in that: In step S2, the resonance feature set includes a resonance peak frequency field and a resonance peak confidence field corresponding to each grid node. The resonance peak frequency field is a numerical value output from the resonance peak recognition neural network model, used to characterize the main resonance frequency corresponding to each grid node. The resonance peak confidence field is a confidence value output from the resonance peak recognition neural network model, used to characterize the stability of the corresponding resonance peak frequency. The resonance peak confidence field is obtained based on a confidence estimation mechanism. This confidence estimation mechanism includes: statistically analyzing the energy change trend of the resonance response feature tensor at continuous frequency sampling points; combining and calculating the energy gradient field, frequency local consistency field, and outlier suppression field near the resonance peak frequency; and inputting these fields into the resonance peak recognition neural network model according to preset rules to form a confidence output field.

5. The intelligent seismic exploration method based on multi-mode fusion according to claim 4, characterized in that: In S3, the multi-mode fusion feature vector is a one-dimensional vector formed by concatenating the resonance feature fields corresponding to each grid node with other observation data fields in the multi-mode observation dataset in a fixed order. It is used to uniformly represent the features of multiple types of observation information at the same grid node. The data structure of the multi-mode fusion feature vector is a set of vectors arranged by grid node index. Each vector is composed of the resonance peak frequency field, the resonance peak confidence field, and other observation data fields in sequence. The construction method of the multi-mode fusion feature vector is as follows: the various observation data fields in the multi-mode observation dataset are aligned according to the grid node index, and the aligned fields are arranged and combined in a fixed field order to form the multi-mode fusion feature vector.

6. The intelligent seismic exploration method based on multi-mode fusion according to claim 5, characterized in that: In S3, the learnable modal weighting mechanism is a set of parameters that sets corresponding modal weight parameters for various observation features in the multi-mode fusion feature vector and updates each modal weight parameter during the inversion process. It is used to adjust the contribution of different observation features in the inversion calculation. The learnable modal weighting mechanism is as follows: in the multi-mode fusion inversion model, a first modal weight parameter is set for the resonance feature field, and a second modal weight parameter and subsequent modal weight parameters are set for other observation feature fields. Each modal weight parameter is combined with the feature value of the corresponding field according to a preset operation rule, and each modal weight parameter is updated based on the inversion error during the inversion model training process.

7. The intelligent seismic exploration method based on multi-mode fusion according to claim 6, characterized in that: In S3, the multi-mode fusion inversion model is an inversion calculation model constructed based on multi-mode fusion feature vectors and a learnable mode weighting mechanism. It is used to perform inversion calculations on the multi-mode fusion feature vectors of each grid node to obtain the wave impedance parameter field and density parameter field corresponding to the grid node. The structure of the multi-mode fusion inversion model is defined as follows: an input layer receives the multi-mode fusion feature vectors, a feature weighting layer with embedded mode weighting parameters performs mode weighting processing on the input vectors, a hidden layer performs feature mapping operations and obtains the relevant feature expressions of physical property parameters, and an output layer generates the wave impedance parameter field and density parameter field. The multi-mode fusion inversion model performs inversion calculations on the multi-mode fusion feature vectors in conjunction with the learnable mode weighting mechanism, specifically as follows: the multi-mode fusion feature vectors are input into the feature weighting layer of the inversion model, each feature field is weighted according to the mode weighting parameters, and then input into the hidden layer to perform inversion mapping, and the wave impedance parameter field and density parameter field are generated in the output layer.

8. The intelligent seismic exploration method based on multi-mode fusion according to claim 7, characterized in that: In S3, the set of physical property parameters is a data set with grid node index as key and corresponding wave impedance parameter field and density parameter field as value, including the wave impedance parameter field and the density parameter field of each grid node.

9. A multi-mode fusion-based intelligent seismic exploration system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the seismic intelligent exploration method based on multi-mode fusion as described in any one of claims 1-8.

Images