Prediction of rock mass parameters in the TBM tunnel based on BP neural network integrated simulated annealing algorithm

• Published in: Tunnelling and underground space technology Vol. 95; p. 103103
• Main Authors: Liu, B., Wang, R., Zhao, G., Guo, X., Wang, Y., Li, J., Wang, S.
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01.01.2020; Elsevier BV
• Subjects: Algorithms; Artificial neural networks; Boring machines; BP neural network; Compressive strength; Computer simulation; Datasets; Lithology; Mathematical models; Model accuracy; Model testing; Neural networks; Optimization; Parameters; Predictions; Rock masses; Rocks; Simulated annealing; TBM; Tunnel construction; Tunnels; Underground construction; BP neural network; TBM; Simulated annealing
• ISSN: 0886-7798; 1878-4364
• DOI: 10.1016/j.tust.2019.103103

•A hybrid algorithm (SA-BPNN) which integrates a back propagation neural network with simulated annealing is developed.•SA-BPNN models are established for predicting rock mass parameters, including UCS, Bi, DPW, and α.•The inputs of these models are TBM driving data, including Th, Tor, PR and CP.•These prediction models are established based on 360 samples collected from the Water Supply Project from Songhua River.•The models predict more accurate results for rock mass parameters than common BPNN method. The prediction of rock mass parameters is of great significance in ensuring the safety and efficiency of tunnel boring machine (TBM) tunnel construction. Previous studies have confirmed the existence of a relationship between TBM driving parameters and rock mass parameters. In this work, we attempt to utilize the TBM driving parameters to predict rock mass parameters, including uniaxial compressive strength (UCS), brittleness index (Bi), distance between plane of weakness (DPW), and the orientation of discontinuities (α). We propose a hybrid algorithm (SA-BPNN) which integrates the back propagation neural network (BPNN) with simulated annealing (SA). A three-layer BPNN model was trained, using TBM driving and rock mass parameters from the Songhua River water conveyance project. We collected 320 samples, and randomly selected 280 of these to train the model, while the remaining 40 samples made up the first dataset to test the model. The predicted mean absolute percentage errors (MAPEs) of α, UCS, DPW, and Bi were 7.7%, 13.9%, 12.9%, and 11.0%, respectively, with the corresponding determination coefficient (R2) of 0.845, 0.737, 0.731, and 0.657, respectively. Another 40 samples with different lithology were collected to verify the model. Although the prediction results were not as good as those from the first dataset, they were still acceptable. The results reveal that the SA-BPNN model has a relatively high accuracy. To verify the optimization effect of the SA method on the BPNN algorithm, a BPNN model was established and tested. The results of the SA-BPNN model were more accurate than those of the BPNN model.