CN-121745410-B - Stratum profile prediction method and system based on boundary index and ensemble learning

CN121745410BCN 121745410 BCN121745410 BCN 121745410BCN-121745410-B

Abstract

The invention provides a stratum profile prediction method and system based on limit indexes and ensemble learning, wherein the method comprises the steps of obtaining a two-dimensional profile discrete sampling point set to be predicted, and extracting space coordinates, natural water content, liquid limit and plastic limit; calculating an Altaibao derivative index based on a liquid limit and a plastic limit, wherein the Altaibao derivative index comprises a plasticity index and a liquid index, constructing a multi-dimensional feature vector by combining a space coordinate and a geotechnical index, constructing a multi-stage weak supervision labeling mechanism by using the Altaibao limit index to generate a weak supervision soil class label, taking the multi-dimensional feature vector as an input and the label as a target to train an integrated learning classification model, gridding a section to be predicted, obtaining a grid node physical index by spatial interpolation, constructing a grid point feature vector, inputting the model to obtain a predicted soil class label, and outputting a two-dimensional stratum distribution matrix. The invention integrates the conventional geotechnical limit index and the space information, and realizes the continuous two-dimensional stratum profile prediction with automation, stabilization and engineering rationality.

Inventors

WANG LIYING
WANG LINGYU
CHEN BIN
WANG KAI
LU XING
ZHANG YANLIN
LI GUOAN
NI PENG
YANG JINZE
ZHANG XULIANG
LIU DONGXU
YANG LIANJIANG
LIU JUNLI
SHEN DI
Qi Kehao
YUAN XIANG
WEI XINGXING
HE GUANG
WANG ZICHAO
BAI DONG
WENG XIAOCHUAN
TU JIE
WANG JIAN
HAN GU

Assignees

中铁十八局集团第五工程有限公司
中铁十八局集团有限公司

Dates

Publication Date: 20260508
Application Date: 20260224

Claims (10)

1. The stratum profile prediction method based on the limit index and the ensemble learning is characterized by comprising the following steps: S1, acquiring a discrete sampling point set in a two-dimensional section to be predicted, and extracting conventional geotechnical test indexes of each sampling point, wherein the conventional geotechnical test indexes at least comprise space coordinates (x, z), natural water content w, liquid limit w L and plastic limit w P ; S2, calculating an Altaibao derivative index based on a liquid limit and a plastic limit, and constructing a multidimensional feature vector together with a space coordinate and a geotechnical index, wherein the Altaibao derivative index at least comprises a plastic index Sum liquid index ; S3, under the condition of lacking a manual layering label, constructing a multi-stage weak supervision labeling mechanism with discrimination priority constraint based on an Altaibao limit index, and generating a weak supervision soil label for model training for a sampling point; s4, training an integrated learning classification model by taking the multidimensional feature vector as input and the weak supervision soil type label as a target; S5, performing gridding treatment on the two-dimensional section to be predicted to form a regular grid, acquiring physical indexes of each grid node by using a spatial interpolation algorithm, and constructing a grid point feature vector; S6, inputting the feature vectors of all grid points into a trained integrated learning classification model to obtain a prediction soil class label of the global grid points, and outputting a two-dimensional stratum distribution matrix as a stratum profile prediction result.
2. The method for predicting a profile of a subterranean formation based on boundary criteria and ensemble learning of claim 1, wherein said plasticity index Index of liquid properties The following relations are satisfied respectively: , , Wherein, the Is made of natural water content, Is the liquid limit of the liquid, and the liquid is the liquid limit, Plastic index for plastic limit Used for representing the width of soil plastic interval and liquidity index For reflecting the nature of soil the degree of softness of the state.
3. The stratigraphic section prediction method based on boundary index and ensemble learning according to claim 1, wherein in step S2, a multidimensional feature vector is constructed, and the multidimensional feature vector is 7 dimensions, and specifically expressed in the following form: ; Wherein: is a spatial feature for expressing hierarchical spatial continuity; Is the basic physical property characteristic; is a derivative enhancement feature for enhancing sensitivity to soil engineering properties and weak bands.
4. The method for predicting a profile of a subterranean formation based on boundary criteria and ensemble learning of claim 1, wherein said multi-stage weakly supervised labeling mechanism comprises the following non-exchangeable discriminant priorities: A first priority, performing flow molding or near-liquid state discrimination on the sampling point based on the liquid index, when the liquid index Judging that the plastic soil flows when the preset weak threshold is met; The second priority is to judge coarse-grained soil class based on a threshold value of liquid limit only for sampling points which are not judged to be in a plastic flow or near-liquid state; And the third priority is to judge the fine-grained soil class based on the plasticity index interval only for the sampling points which pass the liquid limit judgment.
5. The stratigraphic section prediction method based on limit indexes and integrated learning according to claim 1, wherein the integrated learning classification model is a random forest classification model, the random forest classification model generates a sub-training set in a Bootstrap resampling mode, partial characteristics are randomly selected to participate in splitting when nodes of a decision tree are split, and a soil prediction result is output in an integrated voting mode of a multi-decision tree; The output soil type prediction result is shown as follows: ; Wherein, the Representing the predicted continuous value of the value, Is the first A prediction result of the decision tree; Is the number of trees in the forest, and the mode function Representing a majority vote.
6. The stratigraphic section prediction method based on boundary index and ensemble learning according to claim 1, wherein the spatial interpolation algorithm adopts a KNN-based distance neighborhood regression method, and the interpolation formula is: ; Wherein: For grid points Is a result of interpolation of (a); Is the first Index values at adjacent sample points; distance between the grid point and the adjacent sample point; the interpolation field includes: 。
7. The method for predicting a profile of a subterranean formation based on boundary criteria and ensemble learning as claimed in any one of claims 1-6, wherein the prediction occurs during the calculation In the case of=0, the correlation index is subjected to deletion complement or equivalent normalization processing.
8. A stratigraphic section prediction system based on boundary index and ensemble learning, comprising: the data acquisition module is used for acquiring the space coordinates of the discrete sampling points and the conventional geotechnical test indexes; The derived index calculation module is used for calculating a plasticity index and a liquidity index; The feature construction module is used for constructing feature vectors of sampling points or grid nodes; The weak supervision labeling module is used for constructing a multi-stage weak supervision labeling mechanism with priority constraint based on the Altaibao limit index according to a preset non-exchangeable discrimination sequence; the integrated learning modeling module is used for training and outputting soil prediction results; the grid construction and interpolation module is used for generating a two-dimensional continuous parameter field; And the profile output module is used for outputting the two-dimensional stratum distribution matrix.
9. The system of claim 8, wherein the weakly supervised labeling module is configured to allow only sample points that are not determined by the high priority discriminant rules to enter a next labeling phase.
10. The system according to claim 8 or 9, wherein the two-dimensional stratum distribution matrix output by the profile output module corresponds to the regular grid nodes one by one, and is used for generating a graphical display result of stratum profiles.

Description

Stratum profile prediction method and system based on boundary index and ensemble learning Technical Field The invention relates to the technical fields of geotechnical engineering, engineering geology and intelligent investigation, in particular to a stratum profile prediction method and system based on limit indexes and ensemble learning. Background In engineering investigation practice under complex stratum conditions such as soft soil distribution areas, stratum division depends on borehole sampling, indoor geotechnical test and experience discrimination of engineering technicians for a long time. Limited by survey costs, construction period and field conditions, the number of boreholes is generally limited, the spatial arrangement is sparse, and actual formations often exhibit significant heterogeneity and gradient characteristics. The traditional manual layering method is strong in subjectivity and is easily influenced by personal experience differences, different technicians can obtain different layering results under the same data, standardization and multiplexing of results are not facilitated, and the requirements of the current engineering design and numerical analysis on high-precision and automatic stratum models are not met. The Altaibao limit index is the most convenient and widely applied geotechnical test parameter obtained in engineering investigation, but is mainly used for the famous or physical and mechanical property analysis of single-point soil samples at present, and a mature scheme for carrying out two-dimensional stratum profile space distribution inference by effectively utilizing the index under the conditions of limited sample points and lacking definite manual layering labels is not formed. Although the existing stratum identification method based on machine learning has advanced to a certain extent, the existing stratum identification method generally depends on a large amount of manually marked training samples or detailed geological priori information, is difficult to adapt to the situation of 'few samples and weak labels' common in actual engineering, and limits popularization and application of the stratum identification method. In particular, the prior art has the following shortcomings in sites with complex stratum conditions: 1. The method is highly dependent on manual experience, has low automation level, and is difficult to realize standardization and high-efficiency output; 2. The sparse drilling results in single-point classification or the difficulty in accurately describing the space continuous change of a stratum interface by a simple connecting line, and the error is larger at a transition area and a local fluctuation position; 3. the main stream machine learning method needs a large amount of labeling data, and the unified layering labels are difficult to obtain by a system in actual engineering; 4. the spatial interpolation technology can only estimate continuous parameters, can not directly output stratum types, still needs to determine boundaries by means of threshold division or manual interpretation, and has insufficient stability; 5. The existing method has the defects that the physical property indexes or the single dimension of the spatial relationship are emphasized, a unified prediction model capable of simultaneously fusing physical property characteristics and spatial position information of a soil body is lacking, and a reliable stratum profile result is difficult to obtain under the condition of limited data. In view of the foregoing, there is a need for a method and a system for predicting a stratigraphic section based on boundary indexes and ensemble learning, which can comprehensively utilize conventional geoboundary indexes and spatial information under the conditions of few samples and weak labels, realize automatic, stable and engineering-reasonable prediction of continuous two-dimensional stratigraphic sections, and improve investigation efficiency and precision. Disclosure of Invention The invention aims to provide a stratum profile prediction method and system based on limit indexes and ensemble learning, and aims to solve the technical problem that continuous stratum profiles are difficult to automatically infer by using conventional geotechnical indexes under the conditions of few samples and weak labels. In order to achieve the above object, in a first aspect, the present invention provides a method for predicting a formation profile based on boundary index and ensemble learning, including the steps of: the method comprises the following steps: S1, acquiring a discrete sampling point set in a two-dimensional section to be predicted, and extracting conventional geotechnical test indexes of each sampling point, wherein the conventional geotechnical test indexes at least comprise space coordinates (x, z), natural water content w, liquid limit w L and plastic limit w P; S2, calculating an Altaibao derivative index based on a liquid limit