Joint inversion and analysis method of tunnel seismic-electrical-radar based on multimodal learning

Abstract

The present invention belongs to the field of geophysical exploration technology, and provides a tunnel seismic-electric-radar joint inversion and analysis method based on multimodal learning, inverts the detection data of tunnel seismic wave method, electrical method and radar method, and obtains the corresponding wave velocity, resistivity and radar imaging results; for the imaging results of resistivity and wave velocity, a three-dimensional adaptive convolutional network is used to extract gradient features to obtain resistivity features and wave velocity features; for the imaging results of radar, a multi-scale radar gradient data extraction network is used to extract features losslessly to obtain dielectric constant features; an attention mechanism is introduced to adaptively allocate weights to resistivity features, wave velocity features and dielectric constant features, and realize the fusion of gradient features of each parameter; based on the fused features, the final identification result is determined. The present invention effectively integrates the gradient feature information of different scale forms, and provides a data feature fusion basis for deep learning joint inversion.

Description

Technical Field

[0001] The present invention belongs to the technical field of geophysical exploration, and in particular relates to a tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] As the geological conditions of engineering construction become more and more complex, multiple types of disasters such as disaster-causing water bodies, fault fracture zones, strong karst, and weathering deep troughs are coupled and chained, and it is difficult to accurately characterize complex geological bodies with only a single physical parameter. The main reason is that under the complex geological conditions of tunnels, the sources of disasters are highly concealed and multiple types of disasters coexist, resulting in extremely complex spatiotemporal characteristics of geophysical fields such as seismic wave fields, electromagnetic fields, and electric fields, which brings huge challenges to advanced geological forecasting. Once misjudgment or omission, it is easy to induce major geological disasters such as sudden water and mud gushing, and landslides. In the face of high-risk sections under complex geology, it is necessary to use a variety of geophysical exploration methods to conduct comprehensive assessments from multiple angles and dimensions.

[0004] Traditional comprehensive prediction technology based on manual comparative analysis is difficult to fully explore the characteristics of different detection results, and even cannot judge when there are contradictions in the detection results, which brings challenges to comprehensive interpretation and judgment. At the same time, the traditional comprehensive prediction method fails to make full use of the essential characteristics of different geophysical methods, as well as their mutual constraints and data fusion, which makes it difficult to accurately reveal the characteristics and correlation of massive information of multiple methods, affecting the accuracy of the detection results. Deep learning has made significant progress in many aspects such as multimodal data fusion, heterogeneous data fusion, multi-task learning, and external prior knowledge fusion. It can effectively mine and analyze the complex mapping relationship between multimodal heterogeneous input data and multi-task output data, and provide a new solution for the joint inversion of multi-source data, thereby improving the prediction accuracy of tunnel disaster-causing water bodies and structures. However, in the work of tunnel advanced geological prediction, due to the particularity of the tunnel detection environment, the characteristics of seismic wave field, electric field and electromagnetic field data are quite different. The joint inversion method of multi-source data for tunnel advanced geological prediction is in the preliminary research stage and has not yet reached the level of practical application. There are two difficulties in realizing the tunnel joint inversion method based on multimodal learning:

[0005] First, in the application of tunnel seismic data, the small offset distance of the tunnel will cause arc-shaped distortion in its inversion gradient, while the resistivity method is affected by the volume effect and has a diffuse distribution. Therefore, how to extract and fuse network features for different data forms is the first problem that needs to be solved.

[0006] Secondly, the tunnel seismic wave method has a meter-level resolution, while the ground penetrating radar uses broadband high-frequency electromagnetic wave detection, which has a shorter wavelength and sub-meter-level resolution. At the same time, the borehole directional radar can drill deep into the borehole and approach the abnormal body for fine detection. However, due to the limitations of the tunnel observation environment, hardware type and computing resources, two-dimensional detection data profiles are usually used for interpretation in actual engineering. How to extract the high-resolution data features of the ground penetrating radar (drilling directional radar) for detailed characterization and effectively fuse multiple detection data of different dimensions and scale forms is a key issue for multimodal learning fusion networks. In addition, the automation level of existing geophysical prospecting result analysis methods is still low, lacking geological information guidance, and it is difficult to achieve identification and analysis based on the joint inversion results of multi-source information. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a tunnel seismic-electric-radar joint inversion and analysis method based on multimodal learning. The present invention effectively integrates the three-parameter gradient feature information of different scale forms, laying a high-quality data feature fusion foundation for the deep learning joint inversion of complex geological models.

[0008] According to some embodiments, the present invention adopts the following technical solutions:

[0009] A tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning includes the following steps:

[0010] Obtain tunnel seismic wave, electrical and radar detection data;

[0011] Invert the detection data of tunnel seismic wave method, electrical method and radar method to obtain the corresponding wave velocity, resistivity and radar imaging results;

[0012] For the imaging results of resistivity and wave velocity, a three-dimensional adaptive convolutional network is used to extract gradient features to obtain resistivity features and wave velocity features. The learning mode of the three-dimensional adaptive convolutional network adds adaptive adjustment of the weight of the detection direction on the basis of considering the offset position in each direction of space and the network training weight;

[0013] Based on the imaging results of the radar, a multi-scale radar gradient data extraction network is used to perform feature lossless extraction to obtain dielectric constant features. The multi-scale radar gradient data extraction network includes a fully connected and multi-scale convolution structure. The fully connected structure performs a fully connected operation on the number of radar data profiles and the labels of all radar data of each profile. The multi-scale convolution structure introduces hole convolutions with different expansion rates to perform multi-scale feature extraction.

[0014] The attention mechanism is introduced to adaptively assign weights to resistivity features, wave velocity features, and dielectric constant features to achieve gradient feature fusion of each parameter.

[0015] Based on the fused features, the final recognition result is determined.

[0016] As an optional implementation, the learning mode of the three-dimensional adaptive convolutional network is:

[0017] ;

[0018] in, , , are the spatial three-directional offset positions of network node k, c is the number of input channels, b is the bias term, Train the weights for the network, is the input c channel, the position coordinates are ( m,n,l ), m and l are the vertical direction of the tunnel face and the direction of the survey line, and n is the detection direction; is the weight amplitude parameter, and the weight is adaptively adjusted in the detection direction n to perform attenuation compensation.

[0019] As a further limited embodiment, the weight amplitude parameter The initial parameters are set to , is the distance in the gradient image detection direction n to guide the amplitude parameter.

[0020] As an optional implementation, the three-dimensional adaptive convolutional network uses the U-Net inversion network as the skeleton and adds it to the adaptive convolutional network. In the adaptive convolutional network, two three-dimensional adaptive convolution operations are performed, followed by pooling, two-step convolution and residual network convolution. The same operation is repeated twice to convert the pixel data, and then a three-layer three-dimensional deconvolution operation is used to gradually upsample the feature map and restore it to the same size as the input data.

[0021] As an optional implementation, the process of the fully connected structure performing a fully connected operation on the number of radar data profiles and the labels of all radar data of each profile includes: the input data of the fully connected structure is [B, C, L×M+E_c], where B is the batch number of network training, C is the number of radar data profiles, the borehole directional radar generates different two-dimensional profile data according to different angles, L and M are the length and width of the profile, respectively, L×M+ E_c is the label of all radar data of each profile, and E_c is a unit matrix of size C×C, which is used to mark the position information of each profile radar data;

[0022] The fully connected structure performs three fully connected operations on the input data.

[0023] As an optional implementation, the dilated convolution includes three layers, and each layer is provided with a dilation rate parameter, and the dilation rate parameter of each layer increases sequentially.

[0024] As a further limited implementation, the expansion rate parameters of the three layers of dilated convolution are 1, 3 and 5 respectively.

[0025] As an optional implementation method, an attention mechanism is introduced to adaptively allocate weights to resistivity features, wave velocity features, and dielectric constant features. The process of realizing the fusion of gradient features of each parameter includes:

[0026] ;

[0027] Among them, for a given input feature map X , W is the spatial attention weight map, Y To apply the attention weight map to the input feature map and output the feature map, W The calculation is based on the maximum pooling and average pooling Operation, get two feature maps, splice these two feature maps, and input them into a 3D convolutional layer to generate a spatial attention feature map.

[0028] As an optional implementation method, the process of determining the final identification result based on the fused features includes: after generating the spatial attention feature map, splicing it with the geological information vector to form a new feature vector, and inputting the new feature vector combining the spatial attention features and the geological information into a fully connected layer to achieve feature dimensionality reduction and feature selection, and then obtaining the final identification result after being processed by the activation function.

[0029] A tunnel seismic-electrical-radar joint inversion and analysis system based on multimodal learning, comprising:

[0030] A data acquisition module is configured to acquire detection data of tunnel seismic wave method, electrical method and radar method;

[0031] An inversion imaging module is configured to invert the detection data of the tunnel seismic wave method, the electrical method and the radar method to obtain the corresponding wave velocity, resistivity and radar imaging results;

[0032] The three-dimensional adaptive feature extraction module is configured to extract gradient features using a three-dimensional adaptive convolutional network for resistivity and wave velocity imaging results to obtain resistivity features and wave velocity features. The learning mode of the three-dimensional adaptive convolutional network adds adaptive weight adjustment of the detection direction based on the consideration of the offset position in each direction of space and the network training weight.

[0033] The multi-scale feature extraction module is configured to perform lossless feature extraction on the radar imaging result using a multi-scale radar gradient data extraction network to obtain dielectric constant features, wherein the multi-scale radar gradient data extraction network includes a fully connected and multi-scale convolution structure, wherein the fully connected structure performs a fully connected operation on the number of radar data profiles and the labels of all radar data of each profile, and the multi-scale convolution structure introduces hole convolutions with different expansion rates to perform multi-scale feature extraction;

[0034] The feature fusion module is configured to introduce an attention mechanism to adaptively assign weights to resistivity features, wave velocity features, and dielectric constant features to achieve gradient feature fusion of each parameter;

[0035] The identification module is configured to determine a final identification result based on the fused features.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] Aiming at the distorted morphology of seismic inversion gradients in tunnel environments and the volume effect of the electrical method itself, the present invention proposes a dual-channel feature fusion network based on a three-dimensional adaptive convolutional network. The feature capture and fusion of data with different gradient morphologies are achieved through autonomous learning of convolution kernels. Moreover, through the setting of the learning mode, it is possible to realize adaptive weight adjustment in the detection direction, perform attenuation compensation, and guide the amplitude parameters, thereby effectively improving the network's data sensitivity to deep distances.

[0038] Aiming at the two-dimensional distribution morphology and high-resolution characteristics of radar data, the present invention proposes a multi-scale radar gradient data extraction network based on lossless feature extraction. The input information is reorganized and transformed into a three-dimensional volume using a fully connected and multi-scale convolutional structure to ensure a consistent data structure when fused with subsequent deep features of earthquakes and induced polarization. A dilated convolution based on three different expansion rates is then introduced to perform multi-scale feature extraction with a larger receptive field.

[0039] The present invention introduces the attention mechanism and adaptively allocates weights to the three parameters to achieve three-parameter gradient feature fusion of seismograph-radar based on multimodal learning, which effectively solves the challenge of radar data in gradient fusion and improves the sensitivity to multimodal data features of different distribution forms and scales.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0042] Figure 1 This is a flow chart of tunnel seismic-electrical-radar joint inversion based on multimodal learning in this embodiment;

[0043] Figure 2 is a schematic diagram of the resistivity and seismic dual-channel parameter fusion network of this embodiment;

[0044] Figure 3 It is a schematic diagram of the three-parameter characteristic fusion network of resistivity, seismic and radar in this embodiment;

[0045] Figure 4 Schematic diagrams of several typical geological models in the tunnel multi-parameter database established in this embodiment, wherein (a) is a single spherical cave model, (b) is a single ellipsoidal cave model, (c) is a two-cave model, (d) is a single cave and fault nested model, (e) is a two-cave and fault nested model, and (f) is a fissure water body and cave nested model;

[0046] Figure 5 1 is a schematic diagram of the multimodal learning joint inversion results in this embodiment, wherein (a), (d) and (g) are respectively the geological real models of electrical, seismic and radar under complex geological conditions with fault structures and irregular fracture water channels, wherein (b) and (e) are respectively the electrical and seismic inversion results of the joint inversion of seismic and electrical parameters, wherein (c), (f) and (h) are respectively the electrical, seismic and radar inversion results of the joint inversion of resistivity, seismic and radar parameters in this embodiment. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0049] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0050] In the absence of conflict, the embodiments in this application and the features in the embodiments may be combined with each other.

[0051] Embodiment 1

[0052] A tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning, such as Figure 1 As shown, the following steps are included:

[0053] Obtain tunnel seismic wave, electrical and radar detection data;

[0054] Invert the detection data of tunnel seismic wave method, electrical method and radar method to obtain the corresponding wave velocity, resistivity and radar imaging results;

[0055] For the imaging results of resistivity and wave velocity, a three-dimensional adaptive convolutional network is used to extract gradient features to obtain resistivity features and wave velocity features. The learning mode of the three-dimensional adaptive convolutional network adds adaptive adjustment of the weight of the detection direction on the basis of considering the offset position in each direction of space and the network training weight;

[0056] Based on the imaging results of the radar, a multi-scale radar gradient data extraction network is used to perform feature lossless extraction to obtain dielectric constant features. The multi-scale radar gradient data extraction network includes a fully connected and multi-scale convolution structure. The fully connected structure performs a fully connected operation on the number of radar data profiles and the labels of all radar data of each profile. The multi-scale convolution structure introduces hole convolutions with different expansion rates to perform multi-scale feature extraction.

[0057] The attention mechanism is introduced to adaptively assign weights to resistivity features, wave velocity features, and dielectric constant features to achieve gradient feature fusion of each parameter.

[0058] Based on the fused features, the final recognition result is determined.

[0059] Each step is described in detail below.

[0060] The first step is to obtain detection data using tunnel seismic wave, electrical and radar methods.

[0061] In this embodiment, detection data obtained by tunnel seismic wave method, electrical method and radar method are obtained. However, when training each network, simulation data or historical measured data can be used.

[0062] In this embodiment, in order to ensure the performance of deep learning network training, a high-quality data sample library can be established. In order to show the performance of feature fusion of different advanced prediction methods, a cave model is designed for the characteristics of the resistivity method; a model of caves and faults coexisting is designed for the characteristics of the seismic wave method; an irregular fracture water body model is further designed for the characteristics of radar fine detection, and a total of 10,000 sets of model data sample pairs are simulated and generated. For the same geological body, three model assignments of resistivity, wave velocity and radar dielectric constant are generated accordingly. The numerical range of resistivity is set to 0-10000hm, the wave velocity range is set to 0-3000m / s, and the dielectric constant range is set to 4-10.

[0063] Figure 4 Figure 1 is a schematic diagram of several typical geological models in the tunnel multi-parameter database established in this embodiment. Figure 4 The single spherical cave model shown in (a) Figure 4 The single ellipsoid cave model shown in (b) Figure 4 The two cave models shown in (c) are as follows: Figure 4 The nested model of a single cave and fault shown in (d) Figure 4 The nested model of two caves and faults shown in (e) Figure 4 The nested model of fracture water and caves shown in (f).

[0064] in Figure 4 The grey parts in (d), (e) and (f) are faults, and the black parts are caves and fissure water bodies.

[0065] The above model can be used to perform simulation and generate relevant data.

[0066] Of course, in other embodiments, you can also change Figure 4 The models in the model, such as changing the number, size and shape of the cave models, or the type, size, quantity and shape of the models in the nested model.

[0067] In view of the different gradient forms of tunnel resistivity and seismic data, a variable convolution network is used to improve the traditional regular convolution kernel distribution, and a gradient feature extraction method for different tunnel detection data forms based on a three-dimensional adaptive convolution network is proposed, which can effectively realize the feature extraction of tunnel resistivity and wave velocity. In view of the different gradient forms of tunnel resistivity and seismic data imaging results, a spatially variable three-dimensional adaptive convolution network learning model is proposed, and the formula is as follows:

[0068] ;

[0069] in, , , , are the four feature adaptive parameters that the convolutional neural network needs to learn during the training process, c is the number of input channels, b is the bias term, is the network training weight, x is the input c channel, the three-dimensional feature element of the position coordinate (m, n, l), m and l are the vertical direction of the tunnel face and the direction of the survey line, and n is the detection direction. , and is the spatial three-dimensional offset position of the network node k, for example, the kth element of the input centered at (m, n, l), for a 3×3×3 convolution kernel, k is 27; is the weighted amplitude parameter. Due to the attenuation of seismic waves and electric fields with the detection depth, the weight is adaptively adjusted in the detection direction n to compensate for the attenuation. In order to guide the amplitude parameter, the initial parameter of A is set to , The distance of the gradient image detection direction n can effectively improve the network's data sensitivity to deep distances.

[0070] The dual-channel parameter network for tunnel resistivity and seismic is constructed, such as Figure 2 As shown, this embodiment adopts a three-dimensional conventional U-Net inversion network as a skeleton, and constructs an adaptive convolutional network to solve the seismic or resistivity gradient inversion problem.

[0071] The size of the input data is 30×30×30. In order to ensure that non-integer sizes do not appear during the pooling process, the spatial dimensions of the original data are padded and converted to 32×32×32. This preprocessing operation ensures that the network can handle input data of any size. In the adaptive convolutional network, two three-dimensional adaptive convolution operations are first performed, followed by pooling-two-step conventional convolution-residual network convolution. The same operation is repeated twice to convert the original 1 channel containing 32×32×32 pixel data into 512 channels containing 4×4×4 pixel data. This part of the operation becomes the encoder in the convolutional network. In the decoder part, three layers of three-dimensional deconvolution operations are used to gradually upsample the feature map and restore it to the same size as the input data.

[0072] Aiming at the two-dimensional distribution morphology and high-resolution characteristics of radar data, a multi-scale radar gradient data extraction network based on feature lossless extraction is proposed. The network architecture consists of two parts: fully connected and multi-scale convolutional structures. Before the fully connected network is used, the input data is embedded, which helps to reduce the number of model parameters, improve computational efficiency, and better deal with the sparsity problem of high-dimensional data. In the fully connected network learning, the input data is [B, C, L×M+E_c], where B is the batch size of network training (Batchsize), C is the number of radar data profiles, the ground penetrating radar profile is usually 2, and the borehole directional radar can generate different two-dimensional profile data according to different angles. L and M are the length and width of the profile, L×M+ E_c is the label of all radar data for each profile, and E_c is a unit matrix of size C×C, which is used to mark the location information of each profile radar data. Some data formats are as follows:

[0073] ;

[0074] After the information is embedded, the data is fully connected for a total of three times. After each convolution, the number of neurons is 2048, 1024, and 512 respectively. The 512 neuron vectors are reorganized into a 3D volume of 8×8×8. This design is to ensure that the data structure is consistent with the deep feature fusion of subsequent earthquakes and IP. Subsequently, the dilated convolution (also called dilated convolution) based on three different dilation rates is introduced for multi-scale feature extraction.

[0075] The similarity between dilated convolution and ordinary convolution is that the size of the convolution kernel is the same, that is, the number of parameters in the neural network remains unchanged. In addition to the size of the convolution kernel, the dilated convolution also has an expansion rate parameter, which is represented by r in this embodiment. The r of the three-layer dilated convolution is 1, 3, and 5 respectively. Therefore, the experimental dilated convolution has a larger receptive field.

[0076] By introducing the attention mechanism, the weights of the three parameters are adaptively allocated at the decoder end, realizing the three-channel parameter gradient feature fusion of seismic-electrical-radar based on multimodal learning. The calculation formula of the spatial attention mechanism is as follows:

[0077] ;

[0078] Given an input feature map X , W is the spatial attention weight map, Y is the feature map output after the attention weight map is applied to the input feature map. WIt can be seen from the calculation formula that based on the maximum pooling and average pooling operations, two feature maps of size 1×H×W×L can be obtained. Subsequently, the two feature maps are concatenated and input into a 3D convolutional layer to generate a spatial attention feature map of size 1×H×P×L, where H is the image height, P is the image width, and L is the image length.

[0079] The calculation strategy of the channel attention mechanism with Squeeze and Excitation (SE) is as follows:

[0080] (1) Squeeze. In this step, the SE block compresses the feature maps of each channel in the last layer of convolution outputs of seismic wave velocity and resistivity at different encoding stages, as well as the three-scale convolution outputs of radar dielectric constant through global average pooling operation. The feature map of each channel is converted into a scalar value through global average pooling operation, which represents the weight of the global feature of the physical parameter channel. The input three-dimensional gradient feature map is X, and its size is , where C is the number of channels, H is the image height, W is the image width, and L is the image length. The expression formula is as follows:

[0081] ;

[0082] in represents the cth channel in the input feature map, and its spatial coordinates are element, Represents the 1×1×1×C vector Z after compression.

[0083] (2) Excitation. In this step, two requirements need to be met: expressing nonlinear relationships and extracting non-mutually exclusive relationships. The excitation stage needs to have the ability to express nonlinear relationships and use independent fully connected layers to learn the excitation weights of each channel to ensure that the excitation weights of each channel are learned independently. is a weight matrix of C / r×C, which transforms the vector Z from 1×1×1×C to 1×1×1×C / r, followed by a ReLU activation operation. is a weight matrix of C×C / r, followed by a Sigmod activation layer, which outputs a weight vector with the same number of channels W .

[0084] (3) Scale. In the last step, the SE block performs an element-by-element multiplication of the learned weight vector with the original channel feature map.

[0085] ;

[0086] As can be seen from the formula, the Scale stage is to take out each element one by one according to the number of different physical parameter channels and multiply it with the corresponding channel of the feature map. This operation scales the feature map of each channel according to its importance and enhances the representation ability of important channels. After generating the attention feature map, it is spliced ​​with the geological information vector to form a new feature vector, such as Figure 3 As shown, the cat in the figure is the abbreviation of concatenate, which means connection or embedding.

[0087] Next, this vector combining spatial attention features and geological information will be input into a fully connected layer to achieve feature dimensionality reduction and feature selection. After being processed by the activation function, the final identification result Z is output. By incorporating geological information, the network can understand the input features more comprehensively and improve the accuracy and reliability of identification and analysis. Figure 2 The convolution module of the specific feature fusion network is set as shown in Table 1.

[0088] Table 1 Convolution module settings

[0089]

[0090] It should be noted that the Conv3d layer in Table 1 actually contains three parts, namely three-dimensional convolution transformation (Conv3d), normalization (BatchNorm3d) and activation function (the activation function of Block7 is Sigmoid, and the activation function of other modules is ReLU). Bi_out represents the output of Blocki, i.e. the i-th convolution module, i=1,…,6; in addition, the network also includes 2 three-dimensional adaptive convolution layers (ACN3d), 3 three-dimensional maximum pooling operation layers (MaxPool3d) and 3 three-dimensional transposed convolution layers (ConvTranspose3d).

[0091] Figure 5 Schematic diagram of the multimodal learning joint inversion result in this embodiment. Figure 5 Middle (a), Figure 5 Middle (d) and Figure 5 (g) is the geological model form of electrical method, seismic method and radar for complex geological conditions with fault structures and irregular fracture water channels. Figure 5 (b) and Figure 5 (e) From the results of the joint inversion of seismoelectric dual parameters and the electrical and seismic inversion, it can be seen that the shape of the water body cannot be accurately identified, and even the electrical inversion results are incomplete in the characterization of the fault structure. However, after adopting the joint inversion strategy of seismoelectric-radar three parameters, the inversion results have been significantly improved. In the seismoelectric-radar three-parameter inversion results, it can be observed that the effects have been improved in many aspects. First, Figure 5(c) The fault structure morphology in the electrical inversion results has been significantly improved, more accurately depicting the complexity of the underground structure. This shows that the introduction of radar data has played a key role in making up for the shortcomings of the electrical method in the fault area in the seismic-electrical joint inversion. Secondly, for the water body morphology of irregular fracture water channels, Figure 5 The seismic inversion results in (f) also show significant improvement. Figure 5 The radar results in the middle (h) are more accurate in depicting the morphology of water bodies than the previous two.

[0092] This embodiment constructs a spatially variable three-dimensional adaptive convolutional network based on the differences in morphological characteristics of resistivity and wave velocity. The free change of the convolution kernel is realized through a three-dimensional linear interpolation algorithm, breaking the traditional convolution kernel distribution pattern, so that it can better capture the different morphological characteristics of gradient data.

[0093] Based on the dimensionality and resolution differences between the dielectric parameters and wave velocity / resistivity in radar inversion, a lossless feature extraction based on embedded vectorized tags is proposed, and different encoding features with the same receptive field as the seismoelectric are generated through dilated convolution, such as Figure 3 As shown in the figure, the Spatial Attention Module (SAM) allows the network to pay more attention to small-scale regional features when processing radar data;

[0094] After extracting multi-parameter gradient features, we further studied the feature fusion algorithm. We constructed a dual-channel feature fusion network based on the U-net three-dimensional convolutional neural network for seismic and electrical joint inversion. We further integrated the radar parameter gradient features at different scales into the decoder. By introducing the channel attention module of Squeeze and Excitation (SE), the network can reproduce and calibrate the useful channel feature information related to the inversion target, thereby effectively fusing the three-parameter gradient feature information of different scales and morphologies, laying a high-quality data feature fusion foundation for the deep learning joint inversion of complex geological models.

[0095] It should be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of an all-hardware embodiment, an all-software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of one or more computer-usable storage media (including but not limited to disk storage, CD - ROM , optical storage, etc.).

[0096] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0097] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.

Claims

1. A tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning, characterized by: The following steps are involved: Obtain tunnel seismic wave, electrical and radar detection data; Invert the detection data of tunnel seismic wave method, electrical method and radar method to obtain the corresponding wave velocity, resistivity and radar imaging results; For the imaging results of resistivity and wave velocity, a three-dimensional adaptive convolutional network is used to extract gradient features to obtain resistivity features and wave velocity features. The learning mode of the three-dimensional adaptive convolutional network adds adaptive adjustment of the weight of the detection direction on the basis of considering the offset position in each direction of space and the network training weight; Based on the imaging results of the radar, a multi-scale radar gradient data extraction network is used to perform feature lossless extraction to obtain dielectric constant features. The multi-scale radar gradient data extraction network includes a fully connected and multi-scale convolution structure. The fully connected structure performs a fully connected operation on the number of radar data profiles and the labels of all radar data of each profile. The multi-scale convolution structure introduces hole convolutions with different expansion rates to perform multi-scale feature extraction. The attention mechanism is introduced to adaptively assign weights to resistivity features, wave velocity features, and dielectric constant features to achieve gradient feature fusion of each parameter. Determine the final recognition result based on the fused features; The learning mode of the three-dimensional adaptive convolutional network is: ； in, , , are the spatial three-directional offset positions of network node k, c is the number of input channels, b is the bias term, Train the weights for the network, is the input c channel, the position coordinates are m,n,l The three-dimensional characteristic element, m and l are the vertical direction of the tunnel face and the direction of the survey line, and n is the detection direction; is the weight amplitude parameter, and the weight is adaptively adjusted in the detection direction n to perform attenuation compensation.

2. The tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning as claimed in claim 1 is characterized in that: Weight Amplitude Parameter The initial parameters are set to , is the distance in the gradient image detection direction n to guide the amplitude parameter.

3. The tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning as claimed in claim 1 is characterized in that: The three-dimensional adaptive convolutional network uses the U-Net inversion network as the skeleton and adds it to the adaptive convolutional network. In the adaptive convolutional network, two three-dimensional adaptive convolution operations are performed, followed by pooling, two-step convolution and residual network convolution. The same operation is repeated twice to convert the pixel data, and then a three-layer three-dimensional deconvolution operation is used to gradually upsample the feature map and restore it to the same size as the input data.

4. The tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning as claimed in claim 1 is characterized in that: The process of the fully connected structure performing a fully connected operation on the number of radar data profiles and the labels of all radar data of each profile includes: the input data of the fully connected structure is [B, C, L×M+E_c], where B is the number of batches of network training, L and M are the length and width of the profile respectively, C is the number of radar data profiles, the borehole directional radar generates different two-dimensional profile data according to different angles, L×M+ E_c is the label of all radar data of each profile, and E_c is a unit matrix of size C×C, which is used to mark the position information of each profile radar data; The fully connected structure performs three fully connected operations on the input data.

5. The tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning as claimed in claim 1 is characterized in that: The dilated convolution includes three layers, each of which is provided with a dilation rate parameter, and the dilation rate parameter of each layer increases sequentially.

6. The tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning as claimed in claim 5 is characterized by: The dilation rate parameters of the three layers of dilated convolution are 1, 3, and 5 respectively.

7. The tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning as claimed in claim 1 is characterized by introducing The attention mechanism performs adaptive weight allocation on resistivity features, wave velocity features, and dielectric constant features. The process of realizing the fusion of gradient features of each parameter includes: ； Among them, for a given input feature map X , W is the spatial attention weight map, Y To apply the attention weight map to the input feature map and output the feature map, W The calculation is based on the maximum pooling and average pooling Operation, get two feature maps, splice these two feature maps, and input them into a 3D convolutional layer to generate a spatial attention feature map.

8. The tunnel seismic-electrical-radar joint inversion and analysis method based on multimodal learning as claimed in claim 1 is characterized in that: Based on the fused features, the process of determining the final recognition result includes: after generating the spatial attention feature map, concatenating it with the geological information vector to form a new feature vector, and inputting the new feature vector combining the spatial attention features and geological information into a fully connected layer to achieve feature dimensionality reduction and feature selection, and then obtaining the final recognition result after being processed by the activation function.

9. A tunnel seismic-electrical-radar joint inversion and analysis system based on multimodal learning, characterized by: include: A data acquisition module is configured to acquire detection data of tunnel seismic wave method, electrical method and radar method; An inversion imaging module is configured to invert the detection data of the tunnel seismic wave method, the electrical method and the radar method to obtain the corresponding wave velocity, resistivity and radar imaging results; The three-dimensional adaptive feature extraction module is configured to extract gradient features using a three-dimensional adaptive convolutional network for resistivity and wave velocity imaging results to obtain resistivity features and wave velocity features. The learning mode of the three-dimensional adaptive convolutional network adds adaptive weight adjustment of the detection direction based on the consideration of the offset position in each direction of space and the network training weight. The multi-scale feature extraction module is configured to perform lossless feature extraction on the radar imaging result using a multi-scale radar gradient data extraction network to obtain dielectric constant features, wherein the multi-scale radar gradient data extraction network includes a fully connected and multi-scale convolution structure, wherein the fully connected structure performs a fully connected operation on the number of radar data profiles and the labels of all radar data of each profile, and the multi-scale convolution structure introduces hole convolutions with different expansion rates to perform multi-scale feature extraction; The feature fusion module is configured to introduce an attention mechanism to adaptively assign weights to resistivity features, wave velocity features, and dielectric constant features to achieve gradient feature fusion of each parameter; The identification module is configured to determine a final identification result based on the fused features; The learning mode of the three-dimensional adaptive convolutional network is: ； in, , , are the spatial three-directional offset positions of network node k, c is the number of input channels, b is the bias term, Train the weights for the network, is the input c channel, the position coordinates are m,n,l The three-dimensional characteristic element, m and l are the vertical direction of the tunnel face and the direction of the survey line, and n is the detection direction; is the weight amplitude parameter, and the weight is adaptively adjusted in the detection direction n to perform attenuation compensation.

Images