One fundamental requirement for surface and underground rock engineering is the ability to characterise the rock mass. An accurate characterisation of the rock mass allows rock engineers to optimize designs safely and economically. A number of classification systems have been developed in the past 50 years to help characterise rock masses such as RMR, Q system, Rmi, and several others. The challenge in developing rock characterisation schemes is often practicality, i.e. creating systems with relatively simple inputs for the everyday practitioner. A second challenge is the degree of subjectivity often inherent in commonly used classification systems. These challenges are especially exacerbated in weak rock or densely jointed rock masses. This paper introduces a new approach to characterise densely jointed rock mass using artificial neural networks and principal component analysis, based on principles commonly used in the geotechnical engineering of rock fills and waste rock dumps. Thirteen input parameters are dimensionally compressed using principal component analysis. Using a global database with more than 100 case histories, an artificial neural network is developed to conduct the rock mass characterisation. The advantage of this approach is less subjectivity as commonly used engineering (quantitative) terms are put to use. A further advantage is that given that the method is site-data dependent, it allows for ‘fine tuning’ of the model for a specific site.
During the past 50 years, several useful rock characterisation and classification systems have been developed (Barton et al., 1974; Bieniawski et al., 1974; Cummings et al., 1982; Deere et al., 1966; Hoek et al., 1995; Laubscher, 1976; Palmstrom, 1995; Ramamurthy & Arora, 1993). A common challenge in these systems is adopting suitable input qualitative or quantitative descriptors of the rock mass matching that are both useful and practical. For practicality, often boundary conditions or limitations of a characterisation/classification scheme are set based on the intended rock engineering use. In general, schemes intended for communication purposes alone tend to be less rigorous than those targeted towards engineering design or rock support, such as Q-System (Barton et al., 1974), RMR (Bieniawski et al., 1974), or Rmi (Palmstrom, 1995). In these systems, included input parameters reflect the intact rock strength, the strength of joints, block size and groundwater conditions. Despite the usefulness of these schemes, the degree of subjectivity increases with the number of joint sets, or increase in weakness of the rock mass. For poor rock mass conditions, it behooves the rock engineer/practitioner, therefore, to employ either more than one characterisation scheme or pursue alternative methods. Unfortunately, there are very few nontraditional approaches to characterise weak or heavily fractured rock mass (Gokceoglu & Aksoy, 2000; Santi, 2006).
