Competitive learning based approaches to tool-wear identification

The tool-wear identification problem suits neural network solution procedures, as the success of S. Rangwala´s (1988) back-propagation approach revealed. However, back-propagation lacks flexibility, requires fully labeled training sets (supervision), and needs complete retraining if the environment should change. Its limitations suggest an alternative approach via unsupervised methods; specifically, competitive learning. The relationship between competitive learning and clustering and issues unique to the neural approach are described and analyzed. The results of applying a Euclidean variant of competitive learning to actual sensor measurements reveal the practical advantages of the unsupervised system, which include the ability to learn and produce classifications without supervision (fully labeled training sets). Moreover, the unsupervised system can indicate environmental changes, and easily shift in and out of training mode without complete retraining