Performance monitoring of rock engineering systems using neural networks

Paper presented at a meeting on: Artificial intelligence in the minerals sector, held in Nottingham, UK, 20 April 1993 (original title: Rock engineering system performance monitoring using neural networks.) Synthetic and analytical approaches to objective-based design of rock engineering projects are discussed. Rock masses are considered as complex dynamic systems with components that concurrently interact with each other and their environment. The existence of a rock engineering system canon (i.e. a set of systems governing rules) is proposed. Its possible components are highlighted and an application is demonstrated using a "rock mass automaton" designed to simulate dynamic heat distribution in rock. In the example given, the canon implements the basic result from probability theory applied to the thermodynamics of solids. Rock engineering system performance is reduced to the mapping of the output of a rockmass to varying input (design) options, within the state space of the systems model. The Kohonen Self-Organising Map provides an extremely useful analytical tool to reduce the hyper-dimensionality of rock engineering system behaviour. Further, the hyper-dimensional state vector is mapped "quasiconformally" such that the neural map tends to spatially organise similar input state vectors, facilitating interpretation of rock engineering system behaviour (performance monitoring). Work is reported on the development and validation of the automation as a basis for an internal representation of the rock mass bahaviour. The results are reported of simulations compensating for the uncertainty of the exact mechanisms operating within the rock mass, using additive noise with the input state vector in the training stages of the self-organising map implementation. Initial work on adopting Grossberg's Adaptive Resonance Theory neural network architecture for classification of rock engineering systems behaviour is also presented