Robust Learning and Identification of Patterns in Statistical Process Control Charts Using a Hybrid RBF Fuzzy Artmap Neural Network Gerson Tontini, Dr. Eng. Management Sciences Department, Regional University of Blumenau - FURB Rua Ant8nio da Veiga 140 890 10-97 1, Blumenau - SC - Brazil, e-mail: tontini@fhrb.rct-sc.br ABSTRACT Today customers are demanding more diversijed products, with higher quality and shorter delivery times. It has led to the adoption of Flexible Manufacturing Systems (FMg. The quality control of the manufacturing process in FMS is a critical factor, requiring Jlexible and intelligent quality control systems that are capable of autonomous pattem identification. Because of its learning and generalization capabilities, neural networks have good perspectives for this task. One of the most important difficulties in pattern identiflcation with neural networks is the sensibility to the presentation order of the training pattems. This paper presents an hybrid network, RBF Fuzzy-Artmap, capable of on-line incremental learning, 98% less sensible to the presentation order of training patterns than the Fuzzy-Artmap network. Also, this work compares the performance of the RBF Fuzzy- Artmap network with the Fuzzy Artmap network in the identijcation of six diflerent ‘$patterns“ in quality control charts. 1. Introduction The demand for more diversified products with better quality has led to a crescent automation of manufacturing processes and to a dramatic reduction of lot sizes. This automation has caused a growing interest for automated process quality control systems. To be useful in real life, an automated process quality control system needs to: a) process simultaneously several types of information, from different sources, and analyze them in real time; b) adapt itself to changes in the process, leaning with problems and experiences that happened during past operation; c) establish correlation between several variables in a non- controlled situation; d) treat new situations through generalization of past experiences; e) treat non-linear input-output relations; f) help the analyst or operator in the diagnostic of quality problems; g) integrate the control of common and special causes of variation. Past works about pattern identification in Statistical Process Control (SPC) [1][2][3] do not satisfy the needs of an automated quality control system stated above. They were limited in scope and applicability. The needs of an automated quality control system matches some properties of neural networks, malung them good candidates for application in automated quality control systems. Although neural networks have good perspectives for thls application, still some difficulties need to be overcame. One of the most important difficulties in pattern identification with neural networks is their sensibility to the presentation order of the training patterns. This problem is particularly worse in networks with on-line learning capabilities as Fuzzy- -P[41. The presentation order of training patterns does not affect too much the perfomxmce of neural networks with off-line training, such as Radial Basis Function (RBF) networks [5] and Backpropagation Perceptrons, because they first process all training set before malung any adjustment in the network. Although it is an advantage, these networks cannot leam new “knowledge” on-line during their operation. They need to learn the new ‘‘knowledge’’ using the new and all old training samples. It difficult their use in problems where the knowledge is dynamic. So, the development of a network with on-line learning capability that is insensible to the presentation order of training patterns is very important for its successN application in problems such as pattern identification and diagnosis in quality control systems. This paper discusses the leaning at node level of the Fuzzy-Artmap and RBF networks, theoretically showing why the Fuzzy-Artmap network is much more sensible to the presentation order of training patterns than the RBF network. After that, this paper presents a RBF Fuzzy- Artmap network, capable of on-line incremental learning, not sensible to the presentation order of training patterns. It is a hybrid network, formed by a RBF network and a Fuzzy-Artmap network. Finally, this work presents the performance of the Fuzzy-Artmap network, and the hybrid RBF-Fuzzy Artmap Network in the identrfication of patterns of a simulated process. 2. Fuzzy-artmap network 0-7803-4859-1/98 $10.0001998 IEEE 1694