Quantifying particle size and size distribution of mine tailings through deep learning approach of autoencoders

Abstract

A new tailings particle size and size distribution (PSD) quantifying method using Autoencoders is presented to transform multidimensional PSD information into a 1-dimensional code. Firstly, an auto-generation algorithm was designed to produce training data. Secondly, training sample diversity and distribution were analyzed and the PSD code characteristics and model performance were investigated. Finally, out-of-sample testing using 30 kinds of nature tailings and literature data was performed. Results show that the auto-generated samples have sufficient diversity. The model is more accurate and suitable for well-graded tailings, the most common type in engineering practice. The PSD codes for well-graded tailings fall into a narrow-centered range from 0.48 to 1.50. PSD code K-means clustering center for the conventional six tailings categories are 1.296, 1.205, 1.021, 0.712, 0.903, 1.123. The classification results are more specific and accurate. In out-of-sample testing of two applications, the Autoencoders model exhibits good generalization ability.