With the development of the mining industry, the increasing global content of tailings is causing environmental pollution and the use of tailings as a cemented paste backfill (CPB) for underground filling is an effective method of disposal. The application of the backfilling mining method is heavily dependent on the stability of the CPB. However, experimentally evaluating the unconfined compressive strength (UCS) of CPB is time-consuming and expensive. This study discussed the feasibility of using a deep neural network (DNN) model to predict the UCS of CPB. This study established the first global dataset of UCS values of CPB through a literature search and laboratory studies, for the first time. The dataset comprised 986 CPB samples, where each sample included 14 feature values and a target value. The dataset was divided into three parts, i.e., a training set, a validation set, and a testing set, which were used to build, validate, and test the model. The correlation coefficient (R), coefficient of determination, mean absolute error, and mean square error metrics were used for model evaluation. After the hyper-parameter tuning, the number of layers in the DNN model was 4, the learning rate was 0.001, and the dropout rate was 0.25. The optimal DNN model achieved excellent predictive performance on this dataset, with R values on the training, validation, and testing sets of 0.989, 0.964, and 0.967, respectively. The feature importance was analyzed using permutation importance and the Shapley Additive Explanation value, with the cement–tailings ratio found to be the most important feature value with the greatest impact on the model output. This study provides, for the first time, a UCS prediction based on the global CPB dataset, a method that can quickly and accurately obtain the UCS of CPB. The method can save a lot of experimental time and provide a theoretical reference for subsequent environmental management as well as tailings utilization, and promotion of paste filling.
Read full abstract