Soft computing approach for classifying and monitoring multivariate attribute processes

The use of the statistical multivariate control charts in real world applications is rare. The many assumptions and computations that are needed to apply these charts often make application difficult. In this research, a soft computing approach is presented as a practical alternative to the statistical multivariate control charts. The approach consists of three models, which are fuzzy neural model, neuro-fuzzy model, and neural network model.; The objective of the fuzzy neural model is to classify and monitor the overall quality of a product based on a set of correlated quality characteristics that are not measurable on a continuous scale such as color, taste, or smell. The model provides a simple mechanism for determining out-of-control situations. A step-by-step procedure for implementing the fuzzy neural model in real life applications is discussed. A method for constructing membership functions is presented. To demonstrate the aim of this model, an application of grading pecans is employed.; On the other hand, the objective of the neuro-fuzzy model is to classify and monitor the overall quality of a product based on a set of correlated quality characteristics that are measurable on a continuous scale such as weight, dimension, or temperature. A step-by-step procedure for implementing the neuro-fuzzy model in real life applications is presented. An illustrative example based on simulated data is used to demonstrate the purpose of the neuro-fuzzy model.; The third model of the soft computing approach, which is the neural network model, is presented through two stages. In the first stage, the model is presented as an alternative approach to the traditional univariate Shewhart X-bar control charts and other neural networks discussed in the literature for identifying univariate process mean shifts. Then in stage two, the same model is extended to monitor and classify multivariate process mean shifts. The performance of the neural network model either for the univariate or multivariate case is tested in terms of the average run lengths and correct classification percentages using simulation.