Development of an Artificial Neural Network for Real-Time Classification of Cone Penetrometer Strain Gauge Data.

Abstract : This document describes the development of an artificial neural-network-based algorithm for classifying soil behavior type from cone penetrometer strain gauge data. The network input consists of the two standard cone penetration test parameters: the logarithm of cone pressure and the percentage of sleeve friction to cone pressure (friction ratio). Network output is a one-of-n coding of 12 soil classifications. Three- and four-layer backpropagation networks are trained to associate 11,000 data points with the appropriate soil type. The best recall performance is obtained from a four-layer, 2 x 15 x 15 x 12 network with a tested accuracy rate of 98.2%. All classification errors occur at the decision boundaries between class regions. The network was incorporated into the data collection software of the prototype SCAPS vehicle in October 1993. The C source code is included as appendix A.