Piece-Wise Linear Estimation of Mechanical Properties of Materials with Neural Networks

Many real-world problems are concerned with estimation rather than classification. This paper presents an adaptive technique to estimate the mechanical properties of materials from acousto-ultrasonic waveforms. This is done by adapting a piece-wise linear approximation technique to a multi-layered neural network architecture. The piece-wise linear approximation network(PWLAN) finds a set of connected hyperplanes that tit all input vectors as close as possible. A corresponding architecture requires only one hidden layer to estimate any curve as an output pattern. A learning rule for PWLAN is developed and applied to the acousto-ultrasonic data. The efficiency of the PWLAN is compared with that of classical back-propagation network which uses generalized delta rule as a learning algorithm.