Convolutional Neural Network for Control Chart Patterns Recognition

Abstract

Control chart, as a basic tool of statistical process control (SPC), has been widely used in monitoring the fluctuation of variables in manufacturing processes. Unnatural control chart patterns (CCPs) are associated with specific causes which can help practitioners control the manufacturing processes. Over the last several decades, many machine learning methods have been applied to the control chart pattern recognition, and have obtained outstanding results. Convolutional neural networks (CNNs), as a state-of-the-art deep machine learning approach, have been demonstrated remarkable performance in many machine learning fields. In this paper, a novel convolutional neural network was proposed to recognize CCPs. Different from traditional classification methods need hand-crafted features which will need a great engineering effort to improve the classifier accuracy, the proposed method extracts multilevel features from input data automatically. Besides, different kind of basic CPPs and concurrent CCPs (two or more basic CCPs occurring simultaneously) recognition model have been investigated by many researchers, but most models would start from scratch and the previous ones would be abandoned when a new CCPs recognition model was built in the research. To solve this problem, the strategy of transfer learning was used in a finely-tuned previous model to accomplish a new CPPs recognition task. The experiment showed that the proposed CNN model outperformed the support vector machine (SVM) model, and that building a CCPs classifier based on transfer learning has a higher performance than building a classifier from scratch.