An unsupervised learning neural algorithm for identifying process behavior on control charts and a comparison with supervised learning approaches

Abstract The applications of supervised pattern recognition techniques on control charts have shown a substantial improvement in the ability to utilize the information of the chart more effectively than conventional run rules. One major assumption underlying this methodology is that the user has a set of well-defined patterns to detect and a sufficient number of training examples. In practice, however, sufficient training examples may not be readily available, owing either to the inability to simulate these patterns or to the lack of real process data. This paper presents a new approach to detect and identify unnatural patterns on control charts based on the unsupervised self-organizing neural paradigm. The unsupervised methodology is based on ART1 networks. The paper discusses training and testing algorithms to train and test the network using a set of unlabelled natural patterns obtained from the process during normal operation. A comparison is also presented between this unsupervised approach and a major supervised methodology, namely, the statistical learning technique. For the unsupervised methodology, the false alarm rate has substantially improved over that of the supervised methodology, while the rate of identification is higher for the supervised system. The higher rate of identification has been achieved at the cost of providing additional unnatural pattern information to implement the supervised system strategy. © 1997 Elsevier Science Ltd