Geotechnical data analysis: Prediction and modeling using a neural-fuzzy methodology

Girish Agrawal, Purdue University

Abstract

A considerable proportion of geotechnical engineering activity may be termed "pattern matching." Computational neural networks and fuzzy theory are presented, from a unified engineering perspective, as a pattern matching technique for modeling and data analysis in geotechnical engineering. Neural networks and fuzzy systems estimate functions from sample data as do conventional statistical approaches. The major difference is that statistical approaches require that we guess how outputs functionally depend on inputs whereas neural and fuzzy systems do not require that we articulate such a mathematical model. Neural networks recognize ill-defined patterns without an explicit set of rules; they adaptively infer it from sample data. They are adaptive model-free estimators. A brief overview of fuzzy theory is followed by a short discussion covering the basics of computational neural networks. A commonly used neural network model is presented and its operation discussed. Existing data is used to demonstrate the comparative ease with which the network can be trained to perform a variety of tasks involving modeling and prediction. The primary focus is on developing a methodology to evaluate the potential for liquefaction at a site. In conventional terms, assessing the potential for liquefaction is seen as a problem of assigning a given site to one of two classes: "will liquefy;" and "will not liquefy." No provision is made to provide a continuity of classification; a site belongs to one of the two classes with no middle ground. The present work contends that a continuous classification scheme is a more realistic approach to use in liquefaction prediction--indeed in most of geotechnical engineering--and hence the issue is discussed in terms of a "possibility" of liquefaction rather than a "probability" of liquefaction. Two other short application examples are also presented. The first of these is concerned with developing a model for estimation and prediction of drained strength parameters; and the second uses measurements of grain size distribution parameters to demonstrate the potential use of neural networks to fill gaps in measured data. The results are very encouraging and demonstrate the significant potential of this methodology for prediction and modeling in geotechnical engineering. The procedures presented in this work are simple to use, easy to incorporate into routine work, and if regularly used can result in substantial improvement in modeling and forecasting soil behavior.

Degree

Ph.D.

Advisors

Bourdeau, Purdue University.

Subject Area

Civil engineering|Geotechnology|Computer science|Artificial intelligence

COinS