Application of full factorial experiment in designing an ANN-based control chart pattern recognizer

Abstract

An artificial neural network (ANN) based model is a common neurocomputing technique which is effective in performing classification tasks. ANN is also known by other names such as connectionism, parallel distribution processing, natural intelligent systems and machine learning algorithm (Barghash and Santarisi, 2004). In statistical process control (SPC), it has been used in automated recognition of control chart patterns (CCPs) since the last 20 years. A review paper on the ANNs applications in the area of CCPs recognition was published in 1998 (Zorriassatine and Tannock, 1998). In 1990’s, numerous ANN training algorithms such as probabilistic neural network, learning vector quantization, and back-propagation (BPN) (Cheng, 1995; Cheng, 1997; Reddy and Ghost, 1998; Guh and Tannock, 1999; Guh and Hsieh, 1999) were proposed. BPN has become the most effective algorithm, widely used for classifying CCPs (Pham and Sagiroglu, 2001). Generally, most literatures reported that the recognition accuracy of the ANNs recognizers were influenced by several design parameters such as network architecture, patterns behaviour, amount of training patterns, and training algorithm. There were a few researches stated that there was no established theoretical method to determine an optimal ANN architecture. For case by case basis, they were commonly determined empirically (Guh et al., 1999; Gauri and Chakraborty, 2006). However, there were researches used the design of experiment (DOE) in selecting an optimal ANN design. For examples, a resolution IV fractional factorial experiment has been used in analyzing the effects of training parameters (Barghash and Santarisi, 2004) and in selecting significant statistical features (Hassan et al., 2006). Research (Barghash and Santarisi, 2004), however, limited their investigation only to the normal and the shift patterns (i.e. minimum shift, shift range, shift percentage). In this paper, full factorial DOE was utilized in investigating the effects of several parameters to the recognition accuracy of a three layer ANN. Then, an optimal ANN design was proposed. Table 1 shows the relationship of such parameters to the design and performance of an ANN.