NEURAL NETWORK MODELING OF FAILURE SURFACE DEVELOPMENT IN BLUFF DISPLACEMENT STUDIES USING A SIMPLE POLE-AND-CABLE SURVEY SYSTEM

A low-cost pole-and-cable slope displacement monitoring system was described by Chase et al. (2001) and has been successfully used to assess the strain history of a slope for the past ten years. In 2003, extensive well drilling was conducted in order to install in-place inclinometers, vibrating wire piezometers, and dewatering wells at the study site. This drilling, in conjunction with borehole gamma logging, revealed exact positions and locations of stratigraphic contacts and shear surfaces. These data were used to construct slope profiles showing positions of failure surfaces imposed on surface displacement maps obtained from the pole-and-cable monitoring system. The failure surface positions (depth below ground) and surface displacement parameters were used as input data to a multi perceptron computer neural net using the commercial software Neural Works Predict. The input elements were cumulative horizontal and vertical surface displacements, differential survey pole inclinations, and soil type while the output element was depth to failure surface. The network was then trained to predict the shear surface depth. An adaptive gradient learning rule was applied. A sigmoid output layer function was used. Using training data, the neural net gave an R2 value of 0.91. When given test data, the neural net gave an R2 value of 0.86. These results showed the potential of predicting shear surface positions in the subsurface from surface displacement measurements.