Control chart patterns recognition with constrained data

Abstract

Recognition and classification of non-random patterns of manufacturing process data can provide clues to the possible causes that contributed to the product defects. Early detection of abnormal process patterns, particularly in highly precise and rapid automated manufacturing is necessary to avoid wastage and catastrophic failures. Towards this end, various control chart patterns recognition (CCPR) methods have been proposed by researchers. Most of the existing control chart patterns recognizers assumed that data is fully available and complete. However, in reality, process data streams may be constrained due to missing, imbalanced or inadequate data acquisition and measurement problems, erroneous entries and technical failure during data acquisition process. The aim of this study is to investigate and develop an effective recognition scheme capable of handling constrained control chart patterns. Various scenarios of data constraints involving missing rates, missing mechanisms, dataset size and imbalance rate were investigated. The proposed scheme comprises the following key components: (i) characterization of input data stream, (ii) imputation and feature extraction, and (iii) alternative recognition schemes. The proposed scheme was developed and tested to recognize the constrained patterns, namely, random, increasing/decreasing trend, upward/downward shift and cyclic patterns. The effect of design parameters on the recognition performance was examined. The Exponentially-Weighted Moving Average (EWMA) imputation, oversampling and Fuzzy Information Decomposition (FID) were investigated. This research revealed that some constraints in the dataset can eventually change the distribution and violate the normality assumption. The performance of alternative designs was compared by mean square error, percentage of correct recognition, confusion matrix, average run length (ARL), t-test, sensitivity, specificity and G-mean. The results demonstrated that the scheme with an ANNfuzzy recognizer trained using FID-treated constrained patterns significantly reduce false alarms and has better discriminative ability. The proposed scheme was verified and validated through comparative studies with published works. This research can be further extended by investigating an adaptive fuzzy router to assign incoming input data stream to an appropriate scheme that matches complexity in the constrained data streams, amongst others.