CN-122021085-A - Ground stress prediction method based on multi-source data cross-modal fusion

CN122021085ACN 122021085 ACN122021085 ACN 122021085ACN-122021085-A

Abstract

The invention provides a ground stress prediction method based on multi-source data cross-modal fusion, which relates to the technical field of underground stress analysis and comprises the steps of extracting local space image features and different terrain region features from plate reconstructed images and dynamic terrain images, inputting the local space image features and the different terrain region features into a multi-modal feature fusion module, carrying out multi-modal feature fusion in a scale alignment, input mapping and depth separable convolution (DWC) mode to obtain a multi-modal geological feature sequence corresponding to a space-time coordinate sequence point by point, carrying out step-by-step downsampling on the multi-modal geological feature sequence through multi-level average pooling to obtain geological features under different scales, carrying out feature decomposition and information mixing through a multi-scale decomposable mixing module, and finally carrying out prediction on the feature sequences with different scales through a multi-predictor mixing module to finally obtain a ground stress prediction result sequence.

Inventors

SHU PENGCHENG
XIE JIANGTAO
YOU YIXUN
Zhou Jinluan
LIU ZE
Zang Yibo
ZHAO LINTAO
WANG PENGCHENG
ZHOU ZAIZHENG
CHANG CHEN
LI LIANFU
GUO XIAO

Assignees

中国海洋大学

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (3)

1. The ground stress prediction method based on multi-source data cross-modal fusion is characterized by comprising the following steps of: acquiring a point cloud space-time coordinate sequence, a plate reconstruction image sequence and a dynamic topography image sequence; Performing cross-modal feature fusion on the point cloud space-time coordinate sequence, the plate reconstruction image sequence and the dynamic terrain image sequence to obtain a multi-modal geological feature sequence corresponding to the point cloud space-time coordinate sequence point by point; modeling the multi-mode geological feature sequence and predicting the ground stress sequence corresponding to the point cloud space-time coordinate sequence point by point.
2. The method for predicting the ground stress based on the cross-modal fusion of multi-source data according to claim 1, wherein the cross-modal feature fusion is performed on the point cloud space-time coordinate sequence, the plate reconstruction image sequence and the dynamic terrain image sequence to obtain a multi-modal geological feature sequence corresponding to the point cloud space-time coordinate sequence point by point, and the method comprises the following steps: Reconstructing image sequences for slabs The method comprises the steps of respectively extracting local space image features and different terrain region features by using a double-flow network, wherein the specific process is as follows: Reconstructing a sequence of images from a slab Inputting the multilayer perceptron MLP to perform characteristic dimension increasing, wherein the formula is as follows: Wherein, the And As the weight of the material to be weighed, And In order for the offset to be a function of, The function is activated for the purpose of Gelu, The feature vector is obtained after dimension rising; Obtaining a global longitude and latitude grid and a time stamp corresponding to each plate reconstruction image according to the dimension of the plate reconstruction image, positioning point cloud space-time coordinates in the longitude and latitude grid and the time stamp, determining a coordinate point corresponding to each point cloud space-time coordinate in a plate reconstruction image sequence, drawing a circle by taking the coordinate point as a center and r as a radius for one coordinate point, and forming other coordinate points and central coordinate points in a range into a local area feature Local information aggregation is carried out through a single-layer linear mapping, and the formula is as follows: Wherein, the As the weight of the material to be weighed, In order for the offset to be a function of, N is the length of the space-time coordinate sequence of the point cloud, and finally the local space image characteristic sequence corresponding to the space-time coordinates of all the point clouds is obtained ; Image sequence reconstruction of slabs by multi-modal segmentation of large models SegFormer Dividing to obtain divided region feature vectors with different topographic region features Wherein K is the number of the segmented regions; Self-attention mechanism pair Performing multi-source feature interaction to obtain global topographic region features The formula is as follows: Wherein, the , , , 、、 In the form of a matrix that can be learned, In order for the scaling factor to be a factor, Is that A function; for each point in the point cloud space-time coordinate sequence, matching the coordinates to the belonging divided area, and taking the characteristic of the divided area as the topographic area characteristic of the coordinate point Finally, feature sequences of different terrain areas corresponding to all the point cloud space-time coordinates are obtained ; For dynamic topography image sequences The local spatial image characteristic sequence of the dynamic topographic image is obtained by processing the same network structure and flow as the plate reconstruction image processing And different terrain feature sequences ; Space-time coordinate sequence of point cloud Local spatial image feature sequence of plate reconstructed image And different terrain feature sequences Local spatial image feature sequence of dynamic topography images And different terrain feature sequences Splicing in the channel dimension: Wherein, the Splicing the matrixes; mapping the spliced features to a higher dimensional space: Wherein, the As the weight of the material to be weighed, Is biased; The mapped features are input into a DWC module for joint modeling: Wherein the DWC module consists of two parts of channel-by-channel convolution (DEPTHWISE) and point convolution (Pointwise) to finally obtain the multi-mode fusion characteristic sequence 。
3. The method for predicting the ground stress based on multi-source data cross-modal fusion according to claim 1, wherein modeling the multi-modal geological feature sequence and predicting the ground stress sequence corresponding to the point cloud space-time coordinate sequence point by point comprises: Multi-scale feature sequence obtained through multi-level average pooling downsampling Wherein M is the number of different scales; for multi-scale feature sequences Decomposing to obtain overall trend sequences And detail change sequence The formula is as follows: Wherein, the For the purpose of the averaging pooling operation, Is a filling operation; For sequence of detail changes Gradually aggregating and transmitting local change information in a low-layer fine granularity sequence to a high-layer coarse granularity sequence according to the progressive updating from fine to coarse scale, supplementing detail information required by modeling, and carrying out the following formula: Where M represents a multi-scale hierarchical index, m=1 corresponds to the finest granularity feature, m=m corresponds to the coarsest granularity feature, The function is activated for the purpose of Gelu, And As the weight of the material to be weighed, And Is biased; for the overall trend sequence The method comprises the following steps of gradually updating from coarse to fine according to the scale, guiding the representation learning of a low-layer fine granularity sequence from top to bottom by utilizing global trend information contained in a high-layer coarse granularity sequence, enhancing the global consistency and trend modeling capability of the fine granularity sequence, and carrying out the following formula: Where M represents a multi-scale hierarchical index, m=1 corresponds to the finest granularity feature, m=m corresponds to the coarsest granularity feature, The function is activated for the purpose of Gelu, And As the weight of the material to be weighed, And Is biased; Mixing the information to obtain an overall trend sequence And detail change sequence Fusion back to the whole sequence to obtain The formula is as follows: For the following Using a single linear layer directly for each scale of sequence Predicting, and accumulating the predicted results of each scale to obtain a final predicted result The formula is as follows: , 。

Description

Ground stress prediction method based on multi-source data cross-modal fusion Technical Field The invention relates to the technical field of underground stress analysis, in particular to a ground stress prediction method based on multi-source data cross-modal fusion. Background The formation and evolution of the crust stress field are important foundations for understanding earthquake mechanisms, plate construction motions and oil and gas reservoir and geothermal resource distribution. Traditional ground stress research methods are mainly divided into two types, namely numerical simulation based on a physical model and statistical regression based on field observation. The method is characterized in that stress field distribution is solved by establishing an elastic mechanical or viscoelasticity constitutive equation and combining boundary conditions, but the stress field distribution is limited by model simplification and parameter uncertainty, and the method is difficult to realize continuous prediction on a global scale due to the fact that the latter depends on local measurement data such as borehole breakout, hydraulic fracturing and the like. As the application of deep learning in geology expands, researchers began to try to use architectures such as convolutional neural networks for stress field prediction, however, the existing methods still face two key bottlenecks: First, heterogeneous and fusion difficulties of multi-modal geological data. Global dynamic terrain data describes the surface vertical motion in the form of a high resolution grid, while slab reconstruction data describes the horizontal motion in the form of vector trajectories and slab boundary topologies. There are significant differences in data format, space-time basis and physical meaning. The existing machine learning method generally performs independent processing on a single data source or adopts simple characteristic splicing, and cannot fully mine deep coupling relations among multi-mode data, so that systematic and relevance characteristics of geological processes are difficult to capture by a model. Second, limitations of space-time unified evolution modeling. The ground stress field has strong space-time non-stationary characteristic, is not only controlled by the space of regional construction background, but also dynamically evolves along with the processes of plate recombination, magma activity and the like. While traditional time sequence models (such as long and short time memory networks) can handle time sequences, they lack explicit modeling of spatial topological relationships, and spatial prediction models (such as graph neural networks) tend to ignore long-term dependencies of geologic time scales. Therefore, development of a deep learning prediction method capable of cooperatively fusing multi-modal geological data and having space-time unified evolution modeling capability is urgently needed to realize continuous and high-precision reconstruction of a global ground stress field. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides a ground stress prediction method based on multi-source data cross-modal fusion, which solves the problem of multi-modal geological data fusion by introducing a multi-modal segmentation large model and a self-attention mechanism, and combines a ground stress multi-scale space-time hybrid prediction model to perform ground stress prediction. In order to achieve the above purpose, the invention provides a ground stress prediction method based on multi-source data cross-modal fusion, which comprises the following steps: acquiring a point cloud space-time coordinate sequence, a plate reconstruction image sequence and a dynamic topography image sequence; Performing cross-modal feature fusion on the point cloud space-time coordinate sequence, the plate reconstruction image sequence and the dynamic terrain image sequence to obtain a multi-modal geological feature sequence corresponding to the point cloud space-time coordinate sequence point by point; modeling the multi-mode geological feature sequence and predicting the ground stress sequence corresponding to the point cloud space-time coordinate sequence point by point. Optionally, the cross-modal feature fusion is performed on the point cloud space-time coordinate sequence, the plate reconstructed image sequence and the dynamic terrain image sequence to obtain a multi-modal geological feature sequence corresponding to the point cloud space-time coordinate sequence point by point, including: Reconstructing image sequences for slabs The method comprises the steps of respectively extracting local space image features and different terrain region features by using a double-flow network, wherein the specific process is as follows: Reconstructing a sequence of images from a slab Inputting the multilayer perceptron MLP to perform characteristic dimension increasing, wherein the formula is as follows: