INTRODUCTION
Comprehensive classification of the stability of surrounding rock and the blastability and drillability of rock is one of the most complex, knotty prob- lems in rock mechanics design. It aims at the integration of crushing rock (rock drilling and blasting) and preventing rock from being fractured (rein- forcing and stabilizing surrounding rock). Insufficient knowledge of classifi- cation and lack of characterizing information often restrict the existing clas- sification methods from getting optimal results. Input of noisy and incom- plete data may lead to a wrong classification or failure in classification. In these eases, the newly-developed theory,-artificial neural network, is more
suitable
The research described in this paper focuses on constructing a neural net- work classification system for underground opening design aids. Two multi- layer feedforward neural network learning systems used in cascade as com- ponents learn first knowledge from prior case histories and set up nonlinear mapping between various affecting factors and the equivalent categories of rock masses. The knowledge thus learned can let the neural network recog- nizers access previous experience with similar excavations and identify the equivalent categories of rock masses and drilling and blasting requirements that they have not seen previously.
ARTIFICIAL NEURAL NETWORK CLASSIFICATION SYSTEM
An artificial neural network classification system was constructed on SUN 386/33 mhz computer with a math coprocessor. This system consists of eight modulars including collection and input of ease records, ease record database, neural network knowledge base , geological information input, geotechnical database, network learning system , neural network recognizer and result output.
The neural network learning system extracts first complex, hidden and higher order geological features and key geologic parameters and sets up in- put-output pattern pairs from ease histories stored in ease record database. Then, it learns classification knowledge and obtains nonlinear mapping be- tween inputs and outputs using two multilayer feedforward neural networks and a improved back-propagation learning algorithm. We transform network training problem into that of convex programming to improve the perfor- mance of back-propagation learning.
