Back-analysis and reconstruction method of in-situ stress field based on limited sample data

The high cost of in-situ stress measurement limits the number of measuring points. It is difficult to describe and express regional in-situ stress field comprehensively with limited sample measured data. How to accurately con- struct the distribution state of the in-situ stress field within the geological body based on the limited samples of meas- ured in-situ data has always been a key issue of geotechnical engineering concern. Especially with the construction of deep engineering project,the understanding on the distribution law of in-situ stress field is the basis of engineering safety design and disaster prevention work. The GMDH (Group Modeling of Data Handing) neural network algorithm proposed in this paper combined the non-linearity of in-situ stress field distribution with burial depth and the disconti- nuity of in-situ stress field in local geological structure,constructed a boundary condition model for the formation of complex geological body in-situ stress field,and generated complex boundary conditions based on a limited sample of measured data in the field,and fitted the non-linear expression of boundary load. The GMDH neural network algorithm platform was constructed by MATLAB programming,which has the advantages of structural optimization and global op- timization,and it overcame the shortcomings of traditional neural network method such as too many assumptions (net- work structure assumptions) and too simple network structure. It realized the non-linear mapping between the boundary load expression of complex geological body and the stress values of measured points,so as to obtain a more reasonable distribution of in-situ stress field. In order to verify the effectiveness of the method,a two-dimensional geological re- gional model of steeply inclined layer was constructed. Fifteen,twelve,nine and six measuring points were selected for in-situ stress field inversion. The results showed that with the decrease of the number of measuring points,the inversion accuracy of GMDH neural network algorithm was more than 83% . Especially in the limited sample data of six measur- ing points,the inversion accuracy of GMDH neural network algorithm was 84% ,and BP neural network algorithm was 76% . It showed that GMDH neural network algorithm has a higher inversion accuracy in the limited sample data. In addition,the in-situ stress field of Xingshan iron mine was inversely calculated and reconstructed based on limited sample data. The results showed that the inversion accuracy of GMDH neural network algorithm reached 84% ,and the inversion error of stress component of most measuring point was less than 10% . Therefore,GMDH neural network algo- rithm has a good generalization and non-linear data predictability in the inversion calculation of limited sample measur- ing data,which could provide an effective theoretical basis for future increasingly complex engineering design and con- struction.