Control chart pattern recognition using feature-based learning vector quantization

The reported learning vector quantization (LVQ) network-based control chart pattern (CCP) recognizers in literature use raw process data as the input vector and can recognize six basic CCPs only. In this paper, an LVQ network-based CCP recognizer is presented that can recognize eight basic CCPs, using seven extracted shape features from the pattern data as the input vector. The recognition performance of this recognizer is compared with the LVQ network-based recognizer that uses raw process data as the input vector. The results show that the feature-based recognizer results in substantially better recognition performance than the raw data-based recognizer. The confusion matrix reveals that the recognition performance of the feature-based recognizer can be improved further if any feature that is more powerful in discriminating shift and trend pattern can be identified. Comparison of performances of LVQ network-based and multilayer perceptrons (MLP) network-based recognizers (both using extracted features as input vector) reveals that the LVQ network-based recognizer requires much lesser learning time than the MLP network-based recognizer, but results in little inferior recognition performance.