A robust unsupervised consensus control chart pattern recognition framework Siavash Haghtalab a , Petros Xanthopoulos a,⇑ , Kaveh Madani b a Department of Industrial Engineering and Management Systems, University of Central Florida, 4000 Central Florida blvd., Orlando, FL 32816, USA b Centre for Environmental Policy, Imperial College London, 14 Prince’s Gardens, London SW7 1NA, UK article info Article history: Available online 4 May 2015 Keywords: Consensus clustering Unsupervised learning Control chart pattern recognition k-Means Spectral clustering Graph partitioning abstract Early identification and detection of abnormal patterns is vital for a number of applications. In manufac- turing for example, slide shifts and alterations of patterns might be indicative of some production process anomaly, such as machinery malfunction. Usually due to the continuous flow of data, monitoring of man- ufacturing processes and other types of applications requires automated control chart pattern recognition (CCPR) algorithms. Most of the CCPR literature consists of supervised classification algorithms. Fewer studies consider unsupervised versions of the problem. Despite the profound advantage of unsupervised methodology for less manual data labeling their use is limited due to the fact that their performance is not robust enough and might vary significantly from one algorithm to another. In this paper, we propose the use of a consensus clustering framework that takes care of this shortcoming and produces results that are robust with respect to the chosen pool of algorithms. Computational results show that the proposed method achieves not less than 79.10% G-mean with most of test instances achieving higher than 90%. This happens even when in the algorithmic pool are included algorithms with performance less than 15%. To our knowledge, this is the first paper proposing an unsupervised consensus learning approach in CCPR. The proposed approach is promising and provides a new research direction in unsupervised CCPR literature. Ó 2015 Elsevier Ltd. All rights reserved. 1. Introduction Time series analysis is an area of research with numerous appli- cation in many fields of science and engineering (Box, Jenkins, & Reinsel, 2013). In manufacturing, for instance, time series pattern recognition is important since slide alterations might be indicative of a malfunction that requires a course of appropriate corrective actions (e.g. maintenance). Manual monitoring is tedious and requires specialized personnel’s undistracted attention. For this, machine learning based automated algorithms, also known as con- trol chart pattern recognition (CCPR) algorithms, have been pro- posed to detect abnormal behaviors. The term was originally coined by Shewhart (1931). An early taxonomy of the patterns was presented in an early publication of the Western Electric Company (1958). Fig. 1 depicts six of the most common abnormal patterns studied in the literature. These different abnormal patterns are usually related to a specific malfunction and their early detection can provide useful insights for corrective actions and thus improve systems reliability. In the crank case manufacturing operations, up trend and down trend patterns reveal tool wear and malfunction (El-Midany, El-Baz, & Abd-Elwahed, 2010a). Shift patterns might be associated with variation related to operator, material or machine instrument (Davy, Desobry, Gretton, & Doncarli, 2006; El-Midany et al., 2010a). Cyclic patterns are associated with voltage variability (Kawamura, Chuarayapratip, & Haneyoshi, 1988) but they can also appear in manufacturing processes like frozen orange juice packing (Hwarng, 1995). In the car manufacturing industry certain anoma- lies in the automotive body assembly process appear as up/down trends, cyclic, and systematic patterns (Jang, Yang, & Kang, 2003). Up/down trend patterns can be used in order to detect abnormal stamping tonnage signals (Jin & Shi, 2001). Finally up/down trend signals appear in paper making industry (Chinnam, 2002; Cook & Chiu, 1998) whereas uptrend patterns by itself can be used for detecting fault states in end-milling process (Zorriassatine, Al-Habaibeh, Parkin, Jackson, & Coy, 2005). During several years, different pattern recognition algorithms have been studied in the literature with the proposed approaches ranging over a broad spectrum of machine learning algorithms. The majority of the proposed schemes follow the supervised http://dx.doi.org/10.1016/j.eswa.2015.04.069 0957-4174/Ó 2015 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. E-mail addresses: s.haghtalab@gmail.com (S. Haghtalab), petrosx@ucf.edu (P. Xanthopoulos), k.madani@imperial.ac.uk (K. Madani). Expert Systems with Applications 42 (2015) 6767–6776 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa