Predicting compressive strength of sprayed concrete lining in tunnels: Ensemble deep learning with ARF optimization

Abstract

The ability to withstand compression is an important factor in evaluating the effectiveness of Sprayed Concrete Lining (SCL), but the conventional method for determining this characteristic is both time-consuming and physically demanding. This paper presents a new ensemble deep learning (EDL) model for advanced mixture designbased prediction of the compressive strength of SCL (CSSCL). The model consists of five phases: Data acquisition, pre-processing, feature extraction, feature selection, and compressive strength prediction. The collected raw data undergoes pre-processing via data cleaning and transformation. Then, the statistical features such as Principal Component Analysis (PCA), central tendency (“mean, median, mode”), dispersion (“range, second quartile range, variance, and standard deviation”), skewness & coefficient of variation are extracted from the pre-processed data. From the extracted features, the optimal features are selected using the new hybrid optimization model- ArchRatFly Optimization Algorithm (ARF), which is the combination of “Archimedes Optimization Algorithm” (AROA) and “Rat Swarm Optimization Algorithm” (RSA). The compressive strength is then predicted using the EDL model that integrates “Self-Organizing Maps” (SOMs), “Deep Belief Networks” (DBNs), and optimized Autoencoders. In EDL, the SOMs and the DBNs are trained with optimal features. The outcome from SOM and DBN is fed as input to the optimized autoencoder. The final output, compressive strength, is obtained from the optimized autoencoder. The hidden layers of the autoencoder are optimized using the hybrid optimization algorithm AROA and RSA, designed to improve the prediction accuracy of the model. Results in MATLAB show that the proposed model outperforms existing models in performance metrics such as MAE (0.3), MAPE (1.4), RMSE (0.3), MSE (0.1), Correlation coefficient (1.00), and R2 (0.998).