CN-121582714-B - Fracture network multi-step iteration deep learning prediction method for hidden rock mass

CN121582714BCN 121582714 BCN121582714 BCN 121582714BCN-121582714-B

Abstract

The invention discloses a hidden rock mass oriented fracture network multi-step iteration deep learning prediction method, which relates to the technical field of intelligent identification and prediction of geological fracture structures and comprises the steps of extracting fracture network images of a bare rock region, analyzing geometric feature parameters of the fracture network images of the bare rock mass in the bare rock region, dividing the fracture network images of the bare rock mass into different continuous sequence images based on the geometric feature parameters, constructing a training set and a verification set according to the continuous sequence images, constructing a deep learning neural network model, executing a training process by using the training set and the verification set, setting the overlapping rate of different image blocks of the continuous sequence images and adjacent sequences as multiple groups of candidate combinations, generating corresponding training sets and verification sets for each group of candidate combinations, and training and verifying the deep learning neural network model. And carrying out hidden area fracture network prediction by constructing a deep learning neural network model, realizing continuous extrapolation of a masked area fracture structure, and weakening the cumulative effect of prediction errors.

Inventors

JIANG ZHENJIAO
SONG GENFA
ZHA ENSHUANG

Assignees

吉林大学

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (9)

1. A fracture network multi-step iteration deep learning prediction method for a hidden rock mass is characterized by comprising the following steps of, Extracting a fracture network image of an exposed rock area, and analyzing geometric feature parameters of the fracture network image of an exposed rock mass in the exposed rock area; based on the geometric characteristic parameters, the bare rock fracture network image is divided into different continuous sequence images, and the specific steps are as follows, Defining the height and width of the image block according to the crack length and the interval; setting the horizontal overlapping rate and the vertical overlapping rate of the image blocks, and marking the horizontal overlapping rate and the vertical overlapping rate as the horizontal overlapping rate and the vertical overlapping rate respectively; Respectively calculating a horizontal sliding step length and a vertical sliding step length according to the height, the width, the horizontal overlapping rate and the vertical overlapping rate of the image blocks; Moving image blocks from top to bottom in a bare rock mass fracture network image with random positions as a starting point, with vertical sliding step sizes as intervals, moving the image blocks from left to right with horizontal sliding step sizes as intervals, intercepting one image block at each position, sequencing all the image blocks according to the horizontal position order to form a continuous sequence image, and constructing a training set and a verification set according to the continuous sequence image; constructing a deep learning neural network model, and executing a training process by using the training set and the verification set; Setting different image block sizes of continuous sequence images and adjacent sequence overlapping rates as a plurality of groups of candidate combinations, generating a corresponding training set and verification set for each group of candidate combinations, training and verifying a deep learning neural network model, calculating the intersection ratio between a predicted fracture network and a real fracture network, and taking the candidate combination with the largest intersection ratio as a preferred parameter combination; And applying the optimal parameter combination to the iterative reasoning prediction process of the real fracture network training and the deep learning neural network model to generate a hidden rock fracture network prediction result.
2. The hidden rock mass oriented fracture network multi-step iterative deep learning prediction method of claim 1, wherein the extracting of the fracture network image of the bare rock mass and analyzing of the geometric characteristic parameters of the fracture network image of the bare rock mass in the bare rock mass comprises the following specific steps, Taking the field outcrop photo as an exposed rock area fracture network image, carrying out graying and noise filtering on the exposed rock area fracture network image, and generating a pretreated exposed rock area fracture network image; extracting a fracture network image of the bare rock mass in the bare rock zone by using a threshold segmentation, edge detection and deep learning method for the preprocessed fracture network image of the bare rock zone; Based on a skeleton extraction and connected domain marking method, extracting fracture geometric forms piece by piece to form a fracture parameter sample set; and carrying out statistical analysis on the fracture parameter sample set to generate geometric characteristic parameters.
3. The method for multi-step iterative deep learning prediction of a fracture network for a blind rock mass according to claim 2, wherein the geometrical characteristic parameters comprise fracture length, opening degree, trend and spacing.
4. The method for predicting the multi-step iterative deep learning of the fracture network facing the hidden rock mass according to claim 3, wherein the training set and the verification set are constructed according to continuous sequence images, specifically comprising the following steps of, Selecting image blocks at a plurality of continuous positions in the continuous sequence image to generate a fracture image sequence; The fracture image sequence is divided into a training set and a validation set.
5. The method for predicting the multi-step iterative deep learning of the fracture network facing the hidden rock mass according to claim 4, wherein the method for constructing the deep learning neural network model comprises the following specific steps of, Constructing a deep learning neural network model from image to image end; The deep learning neural network model takes an image block as input and outputs a predicted fracture network image with the same size as the input.
6. The method for predicting the multi-step iterative deep learning of the fracture network oriented to the hidden rock mass according to claim 5, wherein the training process is performed by using a training set and a verification set, and the method comprises the following specific steps of, The first crack image block in the same continuous image sequence in the training set As the prediction starting point of the neural network, the second crack image block in the same image sequence is input As a supervision target of the first step, obtaining a first-step prediction output Calculation of And a second crack image block Single step loss between ; Will be As input for the second step prediction of the neural network, the corresponding third fracture image block As a second step supervision target, obtaining a second step prediction output Calculation of And third crack image block Single step loss between ; The single-step loss is calculated by adopting a combination of the Dice loss and the cross entropy loss; the single step loss of each step is weighted and summed according to preset weight, and total loss is generated; and when the parameters are updated each time, back propagation and iterative updating are carried out on the parameters of the deep learning neural network model according to the total loss.
7. The method for multi-step iterative deep learning prediction of a fracture network for a blind rock mass according to claim 6, wherein the training process further comprises, Reasoning is carried out on the verification set after each round of training, the intersection ratio, the precision and the recall rate of each step of prediction output and the corresponding real fracture image block are respectively calculated, the arithmetic average value of all iteration step intersection ratios is calculated, and the arithmetic average value is used as the comprehensive evaluation index of the current step number training; adopting a learning rate cosine annealing adjustment strategy and an early stopping strategy in the training process to enable the deep learning neural network model to terminate training; and selecting the optimal network weight according to the maximum value of the average cross ratio of all iteration steps on the verification set, and taking the optimal network weight as the final model weight of the deep learning neural network model.
8. The method for predicting the multi-step iterative deep learning of the fracture network facing the hidden rock mass according to claim 7, wherein the step of calculating the intersection ratio between the predicted fracture network and the real fracture network, taking the candidate combination with the largest intersection ratio as the preferred parameter combination comprises the following steps of, Selecting and combining various image block sizes, various horizontal overlapping rates, various vertical overlapping rates and model structures, respectively constructing a training set and a verification set according to each combination, and repeatedly executing a training process to generate a plurality of groups of trained deep learning neural network models; performing multi-step iterative prediction on each group of trained deep learning neural network models on the same fracture network image, calculating the intersection ratio, the Dice coefficient, the precision, the recall rate and the pixel accuracy between the predicted fracture network image and the real fracture image block, and recording the corresponding training time and reasoning time; Performing contrast analysis on the evaluation indexes of all parameter combinations, and selecting a preferable image block size, a preferable horizontal overlapping rate, a preferable vertical overlapping rate and a preferable model architecture by taking the average cross ratio of all iteration steps as a target; the preferred image block size, the preferred horizontal overlap rate, the preferred vertical overlap rate, and the preferred model architecture are combined as preferred parameters.
9. The method for multi-step iterative deep learning prediction of the fracture network for the hidden rock mass according to claim 8, wherein the method for multi-step iterative deep learning prediction of the fracture network for the hidden rock mass is characterized by comprising the following steps of, Constructing a sliding window on the bare rock mass fracture network image according to the preferred image block size and the preferred overlapping rate, taking the bare rock mass fracture network image as an initial input area, and taking a hidden rock mass area as an iterative reasoning target area; Extracting a plurality of initial image blocks from the junction of the fracture network image of the bare rock body and the hidden rock body area along the main fracture spreading direction and the vertical direction, taking each initial image block as a sequence first block, and generating a predicted fracture network image by gradually iterating in the horizontal direction and the fracture spreading direction by using a trained deep learning neural network model until the hidden rock body area is covered; and carrying out overlapping fusion on a plurality of predicted fracture network images obtained by different initial positions and different vertical zoning in space, and carrying out weighted average on the predicted probability value of the overlapping region to obtain a hidden rock fracture network prediction result.

Description

Fracture network multi-step iteration deep learning prediction method for hidden rock mass Technical Field The invention relates to the technical field of intelligent identification and prediction of geological fracture structures, in particular to a fracture network multi-step iteration deep learning prediction method for a hidden rock mass. Background The fracture network developed in the natural rock body determines the spatial distribution and evolution characteristics of groundwater, oil gas, geothermal fluid and stress-strain fields, is a key control factor in engineering construction and resource development, traditional fracture characterization depends on geological investigation and limited observation sections, however, fracture structures observed from wells, roadways and outcrops are often discrete, local and easily blocked and limited, obvious incompleteness and discontinuity exist in space, the current research is difficult to reconstruct a continuous and reliable fracture network structure in an invisible area in a two-dimensional and three-dimensional range by utilizing the data, and in recent years, deep learning is introduced into fracture image interpretation by utilizing neural networks such as U-Net, segNet, YOLACT ++, segFormer and the like to carry out semantic segmentation or instance segmentation on CT, drilling core and outcrop images, so that the fracture extraction precision and the automation degree are remarkably improved. The existing method has some defects that in terms of identification and prediction, in-situ fracture characterization which usually stays in observed data, continuous spatial spreading and topological communication relation of the fracture are not considered, particularly in an actual mining area, when a part of area is blocked or not observed, the current fracture in-situ identification mode cannot be reliably extrapolated to a masking area, in terms of fracture extrapolation, the statistical consistency is usually targeted, the generated result randomness is larger, the specific extension path and the fine geometry structure of the actual fracture are difficult to predict, the depicting capability of a circulation channel and a potential fracture surface is limited, special design is not carried out for translational extrapolation of the fracture in space, and the hidden fracture structure cannot be reconstructed in space. Disclosure of Invention The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides a fracture network multi-step iteration deep learning prediction method for the hidden rock mass, which solves the problems that the extrapolation of a shielding area is unreliable and hidden fractures are difficult to reconstruct. In order to solve the technical problems, the invention provides the following technical scheme: the invention provides a hidden rock mass oriented fracture network multi-step iterative deep learning prediction method which comprises the steps of extracting fracture network images of a bare rock region, analyzing geometric feature parameters of the fracture network images of the bare rock mass in the bare rock region, dividing the fracture network images of the bare rock mass into different continuous sequence images based on the geometric feature parameters, constructing a training set and a verification set according to the continuous sequence images, constructing a deep learning neural network model, executing a training process by using the training set and the verification set, setting the overlapping rate of different image blocks of the continuous sequence images and adjacent sequences as a plurality of groups of candidate combinations, generating a corresponding training set and verification set for each group of candidate combinations, training and verifying a deep learning neural network model, calculating the intersection ratio between a predicted fracture network and a real fracture network, taking the candidate combination with the largest intersection ratio as a preferred parameter combination, and applying the preferred parameter combination to the iterative reasoning prediction process of the real fracture network training and the deep learning neural network model to generate a hidden rock mass fracture network prediction result. The invention is a preferable scheme of the hidden rock mass oriented fracture network multi-step iteration deep learning prediction method, wherein the extraction of the fracture network image of the bare rock mass is carried out, the geometric characteristic parameters of the fracture network image of the bare rock mass in the bare rock mass are analyzed, the specific steps are as follows, Taking the field outcrop photo as an exposed rock area fracture network image, carrying out graying and noise filtering on the exposed rock area fracture network image, and generating a pretreated exposed rock area fract