ASA 125th Meeting Ottawa 1993 May

1pSA3. Training and representation issues in machine diagnostic by neural networks.

Samir I. Sayegh

Phys. Dept., Purdue Univ., Fort Wayne, IN 46805

Neural networks have emerged as a viable and effective paradigm for pattern recognition and discrimination. A widely used algorithm, backpropagation, can be sensitive to the representation used for the input patterns to the network. This is particularly true in machine diagnostic. An application is presented to discrimination of good versus bad electric motors. Representations in terms of spectral coefficients, wavelets, and KL coefficients are contrasted.