abstract
- © 2019 Prognostics and Health Management Society. All rights reserved.Since most manufacturing systems generate only a few defects per million of opportunities, rare quality event detection is one of the main applications of the Process Monitoring for Quality philosophy. Single-hidden-layer feed-forward neural networks have been successfully applied to perform this task. However, since the best network structure is not known in advance, many models need to be learned and tested to select a final model with the right number of hidden neurons. A new three-dimensional model selection criterion (3D-NN) is introduced for the application of shallow neural networks to highly/ultra unbalanced binary data structures. Proposed criterion combines three of the most important attributes - prediction, fit, complexity - of a network structure and map them into a three dimensional space to select the best one. It is simple, intuitive and more stable than widely used criteria - Akaike information criterion, Bayesian information criterion and validation cross-entropy error - when dealing with these data structures.