Towards lightweight excavation: Machine learning exploration of rock size distribution prediction after tunnel blasting

Abstract

Advanced and accurate prediction of rock fragmentation distribution can reduce the secondary crushing work, the cost of manual equipment and increase efficiency, thereby enabling tunnel excavation towards lightweighting. To that end, a novel hybrid random forest (RF) model optimized by atomic orbital search (AOS) with Logistic mapping (LM), i.e., LMAOS-RF, was proposed to predict rock size distribution. Five other hybrid models such as epsilon-support vector regression (e-SVR) and nu-support vector regression (n-SVR), backpropagation neural network (BPNN), extreme learning machine (ELM) and kernel extreme learning machine (KELM) optimized by arithmetic optimization algorithm (AOA) were also developed and compared to LMAOS-RF model. Eleven parameters including uniaxial compressive strength (UCS), blasting workface height (H), degree of bonding of bedding surface (DB), total number of holes (TNH), periphery holes spacing (SPH), relief holes spacing (SRH), blasting layer thickness (TH), total charge (TC), concentration ratio (CR), charge structure (CS) and cut holes maximum charge (CHC) were considered to estimate maximum rock size (MRS) for representing rock size distribution. The performances of six hybrid models were evaluated using statistical indices, regression analysis, error analysis and Taylor diagram. The results indicated that the proposed LMAOS-RF model is the best model to explainrelationship between the considered parameters and MRS. Finally, the results of sensitivity analysis demonstrated that TH and SPH are the greatest positive and negative parameter for predicting MRS. Therefore, this paper provides an accurate guidance for field engineers to optimize blasting design to reduce large block rate and further simplify tunneling.