A ground stress prediction method based on multi-source data cross-modal fusion

Abstract

This invention provides a geostress prediction method based on multi-source data cross-modal fusion, relating to the field of underground stress analysis technology. The method includes: extracting local spatial image features and features of different topographic regions from plate reconstruction images and dynamic topographic images; inputting the local spatial image features, features of different topographic regions, and spatiotemporal coordinates into a multimodal feature fusion module; performing multimodal feature fusion through scale alignment, input mapping, and depth-separable convolution (DWC) to obtain a multimodal geological feature sequence corresponding point-by-point to the spatiotemporal coordinate sequence; progressively downsampling the multimodal geological feature sequence through multi-level average pooling to obtain geological features at different scales; then performing feature decomposition and information mixing using a multi-scale decomposable mixing module using two different methods; finally, predicting the feature sequences at different scales through a multi-predictor mixing module to obtain the geostress prediction result sequence.

Description

Technical Field

[0001] This invention relates to the field of underground stress analysis technology, and in particular to a geostress prediction method based on cross-modal fusion of multi-source data. Background Technology

[0002] The formation and evolution of the crustal stress field are fundamental to understanding earthquake mechanisms, plate tectonics, hydrocarbon accumulation, and geothermal resource distribution. Traditional methods for studying geostress mainly fall into two categories: numerical simulation based on physical models and statistical regression based on field observations. The former establishes constitutive equations of elasticity or viscoelasticity and solves for stress field distribution by combining boundary conditions, but it is limited by model simplification and parameter uncertainties; the latter relies on local measurement data such as borehole collapse and hydraulic fracturing, making it difficult to achieve continuous prediction on a global scale.

[0003] As the application of deep learning in geology expands, researchers have begun to explore using architectures such as convolutional neural networks for stress field prediction. However, existing methods still face two key bottlenecks:

[0004] First, there is the challenge of heterogeneity and fusion of multimodal geological data. Global dynamic topographic data depicts vertical surface movement in the form of high-resolution grids, while plate reconstruction data describes horizontal movement using vector trajectories and plate boundary topology. These two types of data differ significantly in data format, spatiotemporal reference, and physical meaning. Existing machine learning methods typically process single data sources independently or employ simple feature stitching, failing to fully explore the deep coupling relationships between multimodal data. This results in models struggling to capture the systematic and interconnected characteristics of geological processes.

[0005] Second, the limitations of unified spatiotemporal evolution modeling. The geostress field has strong spatiotemporal nonstationar characteristics, being both spatially controlled by the regional tectonic background and dynamically evolving with processes such as plate reorganization and magmatic activity. Traditional time series models (such as Long Short-Term Memory networks) can handle time series, but lack explicit modeling of spatial topological relationships; while spatial prediction models (such as neural networks) often ignore the long-term dependencies of geological time scales.

[0006] Therefore, there is an urgent need to develop a deep learning prediction method that can synergistically integrate multimodal geological data and has the ability to model spatiotemporal evolution in a unified manner, so as to achieve continuous and high-precision reconstruction of the global geostress field. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a geostress prediction method based on multi-source data cross-modal fusion. It solves the problem of multimodal geological data fusion by introducing a multimodal segmentation large model and a self-attention mechanism, and combines a multi-scale spatiotemporal hybrid prediction model for geostress prediction.

[0008] To achieve the above objectives, this invention provides a geostress prediction method based on multi-source data cross-modal fusion, comprising the following steps:

[0009] Acquire point cloud spatiotemporal coordinate sequences, plate reconstruction image sequences, and dynamic terrain image sequences;

[0010] Cross-modal feature fusion was performed on point cloud spatiotemporal coordinate sequences, plate reconstruction image sequences, and dynamic topographic image sequences to obtain multimodal geological feature sequences that correspond to each point in the point cloud spatiotemporal coordinate sequences.

[0011] Model the multimodal geological feature sequence and predict the geostress sequence corresponding to each point in the spatiotemporal coordinate sequence of the point cloud.

[0012] Optionally, the cross-modal feature fusion of the point cloud spatiotemporal coordinate sequence, the plate reconstruction image sequence, and the dynamic terrain image sequence to obtain a multimodal geological feature sequence corresponding to each point of the point cloud spatiotemporal coordinate sequence includes:

[0013] For plate reconstruction image sequences A two-stream network is used to extract local spatial image features and features of different terrain regions, respectively. The specific process is as follows:

[0014] Reconstructing image sequences of plates The feature dimensionality is increased by inputting into a Multilayer Perceptron (MLP), as shown in the following formula:

[0015]

[0016] in, and As weight, and For bias, For Gelu activation functions, These are the feature vectors obtained after dimensionality increase;

[0017] Based on the dimensions of the reconstructed plate images, the global latitude and longitude grid and timestamp corresponding to each image are obtained. The spatiotemporal coordinates of the point cloud are located within the latitude and longitude grid and timestamp to determine the coordinate point corresponding to each point cloud spatiotemporal coordinate in the sequence of reconstructed plate images. For a given coordinate point, a circle is drawn with that coordinate point as the center and r as the radius. The other coordinate points within the circle and the center coordinate point together form a local region feature. Local information aggregation is achieved through a single-layer linear mapping, as shown in the following formula:

[0018]

[0019] in, As weight, For bias, N is the length of the point cloud spatiotemporal coordinate sequence. Finally, the local spatial image feature sequence corresponding to all point cloud spatiotemporal coordinates is obtained. ;

[0020] The image sequence of the plate was reconstructed using the large multimodal segmentation model SegFormer. The segmentation process yields feature vectors for segmented regions with different topographic features. , where K is the number of regions after segmentation;

[0021] Self-attention mechanism Multi-source feature interaction is performed to obtain global terrain region features. The formula is as follows:

[0022]

[0023] in, , , , , , For learnable matrices, Scaling factor for function;

[0024] For each point in the spatiotemporal coordinate sequence of the point cloud, the corresponding segmented region is matched to the coordinates, and the features of the segmented region are used as the terrain region features of the coordinate point. Finally, feature sequences of different terrain regions corresponding to all point cloud spatiotemporal coordinates were obtained. ;

[0025] For dynamic terrain image sequences By using the same network structure and process as the image reconstruction process described above, the local spatial image feature sequence of the dynamic terrain image is obtained. and feature sequences of different terrain regions ;

[0026] Spatiotemporal coordinate sequence of point cloud Local spatial image feature sequences of plate reconstructed images and feature sequences of different terrain regions Local spatial image feature sequences of dynamic terrain images and feature sequences of different terrain regions Perform splicing along the channel dimension:

[0027]

[0028] in, For matrix concatenation;

[0029] Map the concatenated features to a higher-dimensional space:

[0030]

[0031] in, As weight, For bias;

[0032] The mapped features are then input into the DWC module for joint modeling.

[0033]

[0034] The DWC module consists of two parts: depthwise convolution and pointwise convolution, ultimately yielding a multimodal fused feature sequence. .

[0035] Optionally, the step of modeling the multimodal geological feature sequence and predicting the geostress sequence corresponding to each point in the spatiotemporal coordinate sequence of the point cloud includes:

[0036] Multi-scale feature sequences are obtained through multi-level average pooling downsampling. , where M is the number of different scales;

[0037] For multi-scale feature sequences Decompose the data to obtain the overall trend sequence. and detailed change sequence The formula is as follows:

[0038]

[0039]

[0040] in, For average pooling operation, For filling operations;

[0041] For detailed change sequences The update is performed layer by layer, from fine to coarse: local change information in the lower-level fine-grained sequence is gradually aggregated and transferred to the higher-level coarse-grained sequence to supplement the detailed information required for modeling, as shown in the following formula:

[0042]

[0043] Where m represents a multi-scale hierarchical index, m=1 corresponds to the finest-grained feature, and m=M corresponds to the coarsest-grained feature. For Gelu activation functions, and As weight, and For bias;

[0044] For the overall trend sequence The update is performed layer by layer, from coarse to fine: Utilizing the global trend information contained in the high-level coarse-grained sequences, the representation learning of the low-level fine-grained sequences is guided from top to bottom, enhancing the global consistency and trend modeling capability of the fine-grained sequences, as shown in the following formula:

[0045]

[0046] Where m represents a multi-scale hierarchical index, m=1 corresponds to the finest-grained feature, and m=M corresponds to the coarsest-grained feature. For Gelu activation functions, and As weight, and For bias;

[0047] The overall trend sequence after information mixing and detailed change sequence By fusing back into the overall sequence, we obtain The formula is as follows:

[0048]

[0049] for A single linear layer is used to directly process the sequence at each scale. Make predictions, and then sum the prediction results at each scale to obtain the final prediction result. The formula is as follows:

[0050] ,

[0051] .

[0052] By adopting the above technical solution, the present invention has at least the following beneficial effects:

[0053] 1. To address the problem that existing technologies in multimodal geological data fusion cannot construct cross-modal causal chains and lack physical consistency in predictions due to underlying heterogeneity, this invention proposes a cross-modal geological semantic consistency fusion module. This module utilizes multimodal segmentation and a self-attention mechanism to extract geological semantic entities from dynamic topographic images and plate reconstruction images. It also models the relationships between different regions based on dynamic weights, thereby achieving semantic alignment of cross-modal features and coupled expression at the physical mechanism level. This mechanism effectively avoids semantic loss caused by simple feature splicing, enabling the model to capture the systematic dynamic relationships of geological processes and significantly improving the physical consistency and interpretability of geostress predictions.

[0054] 2. To address the strong spatiotemporal coupling characteristics in the long-term evolution of the geostress field and the inability of traditional models to simultaneously handle spatial topological constraints and long-term temporal dependencies, this invention proposes a multi-scale spatiotemporal evolution prediction module for geostress. This module achieves joint modeling of small-scale detail fluctuations and large-scale trend evolution through multi-scale sequence decomposition, bidirectional scale mixing, and multi-predictor synergistic enhancement. This structure effectively compensates for the shortcomings of time-series models neglecting spatial topology and spatial models weakening long-term dependencies, ensuring the dynamic consistency and physical rationality of the geostress evolution process. This enables the model to perform continuous and high-precision spatiotemporal reconstruction of the geostress field on a global scale. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A schematic diagram illustrating the principle of a geostress prediction method based on cross-modal fusion of multi-source data provided in this embodiment of the disclosure;

[0057] Figure 2 The results show the predicted geostress state in the boundary region between two oceanic plates. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] The first major drawback of existing technologies stems from the inherent physical nature and structural differences of multimodal geological data. Dynamic topographic data depicts vertical movements driven by mantle convection in a raster format, while plate tectonics reconstruction data describes horizontal plate movements in a vector format. These two types of data are fundamentally heterogeneous in terms of data format, spatiotemporal reference, and dynamic dimensions. This fundamental difference directly leads to traditional machine learning models being limited to shallow strategies such as independent processing or simple feature stitching when fusing data, failing to effectively model cross-modal causal chains. The direct consequence is that models struggle to capture the systematic and interconnected characteristics of geological processes, resulting in predictions lacking physical consistency and mechanistic explanatory power.

[0060] To address the shortcomings of existing technologies in multimodal geological data fusion, the first key innovation of this invention lies in proposing a cross-modal geological semantic consistency fusion module. This module resolves the heterogeneity issues in format, benchmark, and physical meaning between dynamic topography and plate reconstruction image data and point source geostress data. A pre-trained multimodal segmentation model is used to perform geological semantic segmentation on the two types of images, extracting entity features with different geological meanings. Then, a self-attention mechanism is used to dynamically calculate the correlation weights between these cross-modal entity features, while simultaneously fusing spatiotemporal coordinates and local features. This simulates the systematic dynamic processes between different geological regions, enabling the model to capture the causal relationships implicit in geological evolution, significantly improving the physical consistency and interpretability of predictions.

[0061] The second major drawback of existing technologies stems from the inherent characteristics of the geostress field itself: strong spatiotemporal coupling and non-stationary evolution. The geostress field is controlled by spatial tectonic units such as orogenic belts and rift systems, and evolves slowly over hundreds of millions of years with events such as plate reorganization. Furthermore, the dramatic changes in land-sea distribution and coordinate frameworks across different geological eras constitute a fundamental challenge for cross-spatial alignment and unified modeling. This challenge directly leads to structural defects in traditional models: while temporal models such as Long Short-Term Memory (LSTM) networks can capture temporal dependencies, they ignore spatial topology; while spatial models such as Graph Neural Networks can express structural relationships, they weaken long-term evolution, resulting in physical distortions of the reconstructed geostress evolution process.

[0062] To address the limitations of existing technologies in unified spatiotemporal evolution modeling, the second key innovation of this invention lies in proposing a multi-scale spatiotemporal evolution prediction module for geostress. This module introduces a multi-scale spatiotemporal hybrid prediction architecture with residual connections during the prediction stage. Decomposition is applied to multi-scale sequences, and further hybridization is performed in both fine-to-coarse and coarse-to-fine directions, thereby sequentially aggregating small-scale and large-scale information. Then, multiple predictors are further integrated to utilize complementary prediction capabilities in multi-scale observations, thus achieving spatiotemporal collaborative prediction and ultimately reconstructing a geostress field that conforms to physical laws with high accuracy.

[0063] The present invention will be described in detail below with reference to specific embodiments:

[0064] refer to Figure 1 This disclosure provides a geostress prediction method based on cross-modal fusion of multi-source data. The method is implemented based on a constructed prediction model, which consists of two parts: a cross-modal geological semantic consistency fusion module and a geostress multi-scale spatiotemporal evolution prediction module. The method includes the following steps:

[0065] S1. Obtain point cloud spatiotemporal coordinate sequence, plate reconstruction image sequence, and dynamic terrain image sequence.

[0066] In this embodiment, the model input includes three parts: a sequence of reconstructed plate images. Dynamic terrain image sequence and point cloud spatiotemporal coordinate sequence , , ,in Represents longitude. Represents latitude, Represents time, and N is the length of the point cloud spatiotemporal coordinate sequence.

[0067] S2. Perform cross-modal feature fusion on the point cloud spatiotemporal coordinate sequence, plate reconstruction image sequence, and dynamic terrain image sequence to obtain a multimodal geological feature sequence that corresponds to each point in the point cloud spatiotemporal coordinate sequence.

[0068] In this embodiment, the image sequence is reconstructed from the plate. and dynamic terrain image sequence Local spatial image features and features of different terrain regions are extracted. The local spatial image features, features of different terrain regions, and spatiotemporal coordinate sequences of point clouds are input into the multimodal feature fusion module. Multimodal feature fusion is performed through scale alignment, input mapping, and depth separable convolution (DWC) to obtain a multimodal geological feature sequence corresponding to each point of the spatiotemporal coordinate sequence of the point cloud.

[0069] Specifically, this step is achieved through a cross-modal geological semantic consistency fusion module, as follows:

[0070] For plate reconstruction image sequences A two-stream network is used to extract local spatial image features and features of different terrain regions, respectively. The specific process is as follows:

[0071] Local spatial image feature extraction: reconstructing plate image sequences The feature dimensionality is increased by inputting into a Multilayer Perceptron (MLP), as shown in the following formula:

[0072]

[0073] in, and As weight, and For bias, For Gelu activation functions, These are the feature vectors obtained after dimensionality increase;

[0074] Based on the dimensions of the reconstructed plate images, the global latitude and longitude grid and timestamp corresponding to each image are obtained. The spatiotemporal coordinates of the point cloud are located within the latitude and longitude grid and timestamp to determine the coordinate point corresponding to each point cloud spatiotemporal coordinate in the sequence of reconstructed plate images. For a given coordinate point, a circle is drawn with that coordinate point as the center and r as the radius. The other coordinate points within the circle and the center coordinate point together form a local region feature. Local information aggregation is achieved through a single-layer linear mapping, as shown in the following formula:

[0075]

[0076] in, As weight, For bias, N is the length of the point cloud spatiotemporal coordinate sequence. Finally, the local spatial image feature sequence corresponding to all point cloud spatiotemporal coordinates is obtained. ;

[0077] Feature extraction for different terrain regions: SegFormer multimodal segmentation model for reconstructing plate image sequences The segmentation process yields feature vectors for segmented regions with different topographic features. , where K is the number of regions after segmentation;

[0078] Self-attention mechanism Multi-source feature interaction is performed to obtain global terrain region features. The formula is as follows:

[0079]

[0080] in, , , , , , For learnable matrices, Scaling factor for function;

[0081] For each point in the spatiotemporal coordinate sequence of the point cloud, the corresponding segmented region is matched to the coordinates, and the features of the segmented region are used as the terrain region features of the coordinate point. Finally, feature sequences of different terrain regions corresponding to all point cloud spatiotemporal coordinates were obtained. ;

[0082] For dynamic terrain image sequences By using the same network structure and process as the image reconstruction process described above, the local spatial image feature sequence of the dynamic terrain image is obtained. and feature sequences of different terrain regions ;

[0083] Multimodal feature fusion: combining point cloud spatiotemporal coordinate sequences Local spatial image feature sequences of plate reconstructed images and feature sequences of different terrain regions Local spatial image feature sequences of dynamic terrain images and feature sequences of different terrain regions Perform splicing along the channel dimension:

[0084]

[0085] in, For matrix concatenation;

[0086] Map the concatenated features to a higher-dimensional space:

[0087]

[0088] in, As weight, For bias;

[0089] The mapped features are then input into the DWC module for joint modeling.

[0090]

[0091] The DWC module consists of two parts: depthwise convolution and pointwise convolution, ultimately yielding a multimodal fused feature sequence. .

[0092] S3. Model the multimodal geological feature sequence and predict the geostress sequence corresponding to each point of the point cloud spatiotemporal coordinate sequence.

[0093] In this embodiment, the multimodal geological feature sequence is first downsampled stepwise through multi-level average pooling to obtain geological features at different scales. Then, the feature decomposition and information mixing are performed in two different ways through a multi-scale decomposable mixing module. Finally, the feature sequences at different scales are predicted through a multi-predictor mixing module to obtain the geostress prediction result sequence.

[0094] Specifically, this step is achieved through the multi-scale spatiotemporal evolution prediction module for geostress, as follows:

[0095] Multi-scale feature sequences are obtained through multi-level average pooling downsampling. , where M is the number of different scales;

[0096] Multi-scale decomposable hybrid modules for multi-scale feature sequences Decompose the data to obtain the overall trend sequence. and detailed change sequence The formula is as follows:

[0097]

[0098]

[0099] in, For average pooling operation, For filling operations;

[0100] For detailed change sequences The update is performed layer by layer, from fine to coarse: local change information in the lower-level fine-grained sequence is gradually aggregated and transferred to the higher-level coarse-grained sequence to supplement the detailed information required for modeling, as shown in the following formula:

[0101]

[0102] Where m represents a multi-scale hierarchical index, m=1 corresponds to the finest-grained feature, and m=M corresponds to the coarsest-grained feature. For Gelu activation functions, and As weight, and For bias;

[0103] For the overall trend sequence The update is performed layer by layer, from coarse to fine: Utilizing the global trend information contained in the high-level coarse-grained sequences, the representation learning of the low-level fine-grained sequences is guided from top to bottom, enhancing the global consistency and trend modeling capability of the fine-grained sequences, as shown in the following formula:

[0104]

[0105] Where m represents a multi-scale hierarchical index, m=1 corresponds to the finest-grained feature, and m=M corresponds to the coarsest-grained feature. For Gelu activation functions, and As weight, and For bias;

[0106] The overall trend sequence after information mixing and detailed change sequence By fusing back into the overall sequence, we obtain The formula is as follows:

[0107]

[0108] for The multi-predictor hybrid module uses a single-layer linear layer to directly process the sequence at each scale. Make predictions, and then sum the prediction results at each scale to obtain the final prediction result. The formula is as follows:

[0109] ,

[0110] .

[0111] The method described in the above embodiments is used to predict the geostress state in the boundary region between two oceanic plates. This region is located in a complex tectonic zone with multiple plate interactions and has significant plate boundary activity characteristics.

[0112] Based on plate tectonics reconstruction image sequences and dynamic topographic image sequences, combined with point cloud spatiotemporal coordinate data of the study area, cross-modal feature fusion is performed to predict and reconstruct the regional geostress field. Figure 2 The figure shows the predicted results of the geostress state in the boundary region of two oceanic plates. The short lines of different colors and directions in the figure represent the principal stress directions and their relative intensity distribution characteristics at different locations. Red, green and blue correspond to different types or scales of stress response, respectively, and black lines represent the regional tectonic outline or plate boundary.

[0113] from Figure 2 It can be seen that near the western edge of Ocean 1 and the western boundary of Ocean 2, the stress direction exhibits a clear characteristic of distribution along the plate boundary, with local areas showing stress concentration and abrupt changes in direction, reflecting the significant control of plate subduction and collision processes on the regional stress field. Meanwhile, within the plate interior, the stress distribution is relatively continuous and changes gradually, demonstrating strong regional tectonic consistency.

[0114] The results show that the method of the present invention can effectively integrate multi-source geological information, accurately characterize the spatial distribution and evolution of geostress under complex tectonic backgrounds, and has better physical consistency and spatial continuity compared with traditional single data-driven methods.

[0115] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A geostress prediction method based on multi-source data cross-modal fusion, characterized in that, Includes the following steps: Acquire point cloud spatiotemporal coordinate sequences, plate reconstruction image sequences, and dynamic terrain image sequences; Cross-modal feature fusion was performed on point cloud spatiotemporal coordinate sequences, plate reconstruction image sequences, and dynamic topographic image sequences to obtain multimodal geological feature sequences that correspond to each point in the point cloud spatiotemporal coordinate sequences. Model the multimodal geological feature sequence and predict the geostress sequence corresponding to each point in the spatiotemporal coordinate sequence of the point cloud; The cross-modal feature fusion of point cloud spatiotemporal coordinate sequences, plate reconstruction image sequences, and dynamic terrain image sequences yields a multimodal geological feature sequence corresponding to each point in the point cloud spatiotemporal coordinate sequence, including: For plate reconstruction image sequences A two-stream network is used to extract local spatial image features and features of different terrain regions, respectively. The specific process is as follows: Reconstructing image sequences of plates The feature dimensionality is increased by inputting into a multilayer perceptron (MLP), as shown in the following formula: in, and As weight, and For bias, For Gelu activation functions, These are the feature vectors obtained after dimensionality increase; Based on the dimensions of the reconstructed plate images, the global latitude and longitude grid and timestamp corresponding to each image are obtained. The spatiotemporal coordinates of the point cloud are located within the latitude and longitude grid and timestamp to determine the coordinate point corresponding to each point cloud spatiotemporal coordinate in the sequence of reconstructed plate images. For a given coordinate point, a circle is drawn with that coordinate point as the center and r as the radius. The other coordinate points within the circle and the center coordinate point together form a local region feature. Local information aggregation is achieved through a single-layer linear mapping, as shown in the following formula: in, As weight, For bias, N is the length of the point cloud spatiotemporal coordinate sequence. Finally, the local spatial image feature sequence corresponding to all point cloud spatiotemporal coordinates is obtained. ; The image sequence of the plate was reconstructed using the large multimodal segmentation model SegFormer. The segmentation process yields feature vectors for segmented regions with different topographic features. , where K is the number of regions after segmentation; Self-attention mechanism Multi-source feature interaction is performed to obtain global terrain region features. The formula is as follows: in, , , , , , For learnable matrices, Scaling factor for function; For each point in the spatiotemporal coordinate sequence of the point cloud, the corresponding segmented region is matched to the coordinates, and the features of the segmented region are used as the terrain region features of the coordinate point. Finally, feature sequences of different terrain regions corresponding to all point cloud spatiotemporal coordinates were obtained. ; For dynamic terrain image sequences By using the same network structure and process as the image reconstruction process described above, the local spatial image feature sequence of the dynamic terrain image is obtained. and feature sequences of different terrain regions ; Spatiotemporal coordinate sequence of point cloud Local spatial image feature sequences of plate reconstructed images and feature sequences of different terrain regions Local spatial image feature sequences of dynamic terrain images and feature sequences of different terrain regions Perform splicing along the channel dimension: in, For matrix concatenation; Map the concatenated features to a higher-dimensional space: in, As weight, For bias; The mapped features are then input into the DWC module for joint modeling. The DWC module consists of two parts: depthwise convolution and pointwise convolution, ultimately yielding a multimodal fused feature sequence. .

2. The geostress prediction method based on multi-source data cross-modal fusion according to claim 1, characterized in that, The process of modeling multimodal geological feature sequences and predicting the geostress sequence corresponding to each point in the spatiotemporal coordinate sequence of the point cloud includes: Multi-scale feature sequences are obtained through multi-level average pooling downsampling. , where M is the number of different scales; For multi-scale feature sequences Decompose the data to obtain the overall trend sequence. and detailed change sequence The formula is as follows: in, For average pooling operation, For filling operations; For detailed change sequences The update is performed layer by layer, from fine to coarse: local change information in the lower-level fine-grained sequence is gradually aggregated and transferred to the higher-level coarse-grained sequence to supplement the detailed information required for modeling, as shown in the following formula: Where m represents a multi-scale hierarchical index, m=1 corresponds to the finest-grained feature, and m=M corresponds to the coarsest-grained feature. For Gelu activation functions, and As weight, and For bias; For the overall trend sequence The update is performed layer by layer, from coarse to fine: Utilizing the global trend information contained in the high-level coarse-grained sequences, the representation learning of the low-level fine-grained sequences is guided from top to bottom, enhancing the global consistency and trend modeling capability of the fine-grained sequences, as shown in the following formula: Where m represents a multi-scale hierarchical index, m=1 corresponds to the finest-grained feature, and m=M corresponds to the coarsest-grained feature. For Gelu activation functions, and As weight, and For bias; The overall trend sequence after information mixing and detailed change sequence By fusing back into the overall sequence, we obtain The formula is as follows: for A single linear layer is used to directly process the sequence at each scale. Make predictions, and then sum the prediction results at each scale to obtain the final prediction result. The formula is as follows: , 。

Images