Design of an Artificial Neural Network Pattern Recognition Scheme Using Full Factorial Experiment

Automated recognition of process variation patterns using an artificial neural network (ANN) model classifier is a useful technique for multivariate quality control. Proper design of the classifier is critical for achieving effective recognition performance (RP). The existing classifiers were mainly designed empirically. In this research, full factorial design of experiment was utilized for investigating the effect of four design parameters, i.e., recognition window size, training data amount, training data quality and hidden neuron amount. The pattern recognition study focuses on bivariate correlated process mean shifts for cross correlation function, ρ = 0.1 ~ 0.9 and mean shifts, μ = ± 0.75 ~ 3.00 standard deviations. Raw data was used as input representation for a generalized model ANN classifier. The findings suggested that: (i) the best performance for each pattern could be achieved by setting different design parameters through specific classifiers, which (ii) gave superior result (average RP = 98.85%) compared to an empirical design (average RP = 96.5%). This research has provided a new perspective in designing ANN pattern recognition scheme in the field of statistical process control.