| Abstract: | A pattern recognition approach to liquefacation evaluation is propoesed. The state of any soil layer at a level ground site subject to seismic loads is represented by a pattern in a seven-dimensional feature space and can be classified into one of three classes: liquefiable cohesive soil, and non-liquefiable cohesionless soil. The liquefaction potential of the soil layer can be assessed according to the probabilities of the pattern belonging to the three classes. Training patterns derived from field data (piezocone (CPTU) data and maximum ground acceleration) from sites which liquefied or did not liquefy during earthquakes in New Zealand are randomly chosen to design a pattern recognition system to provide an optimal estimation of the liquefaction potential of any soil stratum of interest. Two recognition systems have been set up to estimate the state-conditional probability density function. One is based on a Parzen window approach in which no knowledge of the probabilistic structure of the training patterns is assumed; the other is based on a parameter estimation approach assuming a multivariate normal distribution. The error rate of recognition by the Parzen window approach is 6·9% when taking the window size as 1·5, and the error rate by the parameter estimation approach, which can be easily, is 7·7%. implemented without reference to our training patterns |
