A control chart pattern recognition system for feedback-control processes

• Published in: Expert systems with applications Vol. 138; p. 112826
• Main Authors: De la Torre-Gutiérrez, Héctor, Pham, DucTruong
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 30.12.2019; Elsevier BV
• Subjects: Algorithms; Automatic control; Control chart pattern recognition; Control charts; Control systems; EPC; Fault detection; Fault diagnosis; Feedback; Machine learning; Pattern recognition; Pattern recognition systems; PID control; Process controls; Proportional integral derivative; Radial basis function; Signal monitoring; SPC; Support Vector Machine; Control chart pattern recognition; EPC; PID control; Support Vector Machine; SPC
• ISSN: 0957-4174; 1873-6793; 1873-6793
• DOI: 10.1016/j.eswa.2019.112826

•Automatic control chart pattern recognition (CCPR) in SPC-EPC processes was studied.•CCPR systems were compared for two different EPC controllers.•The best configuration of input factors of the CCPR systems was obtained.•Seven different control chart patterns were identified.•The proposed CCPR system was proved on real world data. The automated diagnosis of control charts to detect faults is a problem studied by many researchers. In recent years, they have turned their attention to processes that do not fulfil the condition of having normally, identically and independently distributed (NIID) variables. With those processes, it is common to have one or more manipulatable variables that can affect the quality characteristic under investigation. The Engineering Process Control (EPC) approach is often used to minimise the variance around the target value of the monitored characteristic by adjusting the manipulatable variables. In this work, a control chart pattern recognition (CCPR) system was developed for processes adjusted by EPC (also known as SPC-EPC or feedback-control processes). This issue of identification of simple control chart patterns for feedback-control processes had previously not been studied. A Machine Learning algorithm was proposed to train a pattern recognition system. All the possible combinations of factors of the CCPR system were studied to determine the combination yielding the highest recognition accuracy, namely, using raw data as input, generating patterns with significance level α = 0.01, monitoring the output signal, and employing a Proportional Integrative Derivative (PID) controller and the Radial Basis Function (RBF) kernel. This combination yielded overall accuracies of 94.18% and 94.14% for the AR(1) and ARMA(1,1) models, respectively.