Compressive strength prediction of sprayed concrete lining in tunnel engineering using hybrid machine learning techniques

Abstract

Sprayed concrete lining is a commonly employed support measure in tunnel engineering, which plays an important role in construction safety. Compressive strength is a key performance indicator of sprayed concrete lining, and the traditional measuring method is time-consuming and laborious. This paper proposes various hybrid machine learning algorithms to accomplish the advanced prediction of compressive strength of sprayed concrete lining based on the mixture design. Two hundred and five sets of experimental data were collected from a water conveyance tunnel in northwestern China for model construction, and each set of data was made up of six basic input variables (i.e., water, cement, mineral powder, superplasticizer, coarse aggregate, and fine aggregate) and one output variable (i.e., compressive strength). In order to eliminate the correlation between input variables, a new composite indicator (i.e., the water-binder ratio) was introduced to achieve dimensionality reduction. After that, four hybrid models in total were built, namely BPNN-QPSO, SVR-QPSO, ELM-QPSO, and RF-QPSO, where the hyper-parameters of BPNN, SVR, ELM, and RF were auto-tuned by QPSO. Engineering application results indicated that RF-QPSO achieved the lowest mean absolute percentage error (MAPE) of 3.47% and root mean square error (RMSE) of 1.30 and the highest determination coefficient (R2) of 0.93 in the four hybrid models. Moreover, RF-QPSO had the shortest running time of 0.15 s, followed by SVR-QPSO (0.18 s), ELM-QPSO (1.19 s), and BPNN-QPSO (1.58 s). Compared with BPNN-QPSO, SVR-QPSO, and ELM-QPSO, RF-QPSO performed the most superior performance in terms of both prediction accuracy and running speed. Finally, the importance of input variables on the model performance was quantitatively evaluated, further enhancing the interpretability of RF-QPSO.