This thesis is intended to apply artificial neural network models to different rock engineering problems, such as the determination of aggregate quality parameters, rock indentation depth, ore boundaries and ore grades. The thesis deals with attempts at predictions of these various features from factors that are known or assumed to have an influence on them using artificial neural networks. A computer program for implementing the Kalman learning algorithm in multilayer neural networks is presented. The thesis comprises a summary and five appended papers. 1. The relationships between a number of aggregate quality parameters, impact value (SPR), abrasion value I (SL) and abrasion value II (KK), and some factors known to influence them, that is, density (Den), point load (Ir), mineral content of quarts (Qz) and mineral contents of brittle material (spr) were analysed on the basis of material from south-west Sweden using neural networks. The prediction accuracies for these three quality parameters were found to be 94.5, 93.5 and 87.9%, respectively, in a test set. 2. On the basis of laboratory tests, neural networks were applied to predictions of rock indentation depths for three different indenter types, cylindrical, hemispherical and truncated, respectively, from a set of known influencing factors, indentation load, density, compressive strength, critical energy release rate, Young's modulus, and Poisson ratio. The average prediction errors for these three different indenters are 10.7, 33.5 and 28.6%, respectively, in a test set.
Luleå: Luleå tekniska universitet, 1997. , 70 p.