Classifying defect factors in fabric production via DIFACONN-miner: A case study

Abstract

In this paper a data mining based case study is carried out in a major textile company in Turkey in order to classify and analyze the defect factors in their fabric production process. It is aimed to understand the causes of the defects in order to minimize their occurrence. The main motivation behind this study is to minimize scrap loses in the company and enabling more sustainable production via data mining. In the analyses, a data mining tool (DIFACONN-miner) that was recently developed by authors is employed. DIFACONN-miner is a novel data mining tool which combines several metaheuristics and artificial neural networks intelligently and it is capable of producing comprehensive classification rules from any data type.

Introduction

Employing data mining in analyzing manufacturing data can be a great support to sustainable and green manufacturing. This is mainly due to fact that knowledge gathered via data mining (DM) approach can be used to improve manufacturing productivity, reduce scrap rates and minimize reworking. All these improvements mean less material and energy consumption which is a prerequisite for sustainable and green manufacturing practices. Based on this motivation, it is decided to carry out a data mining based study for defect diagnosis in fabric production in a major textile company in Turkey. Several other authors’ from the literature have also recognized that data mining can be a very useful and enabling tool for sustainable manufacturing. For example; Modapothala and Issac (2009) employed data mining for analyzing corporate environmental reports. They conclude that their findings confirm reliable direction in sustainability with the use of data mining techniques. Marwah, Sharma, Ramakrishnan, Bash, and Patel (2009a) pointed out that considering the large amount of data produced by physical ecosystems, manual inspection is virtually impossible and thus automated knowledge discovery and data mining techniques are required to synthesize models for enabling sustainable end-to-end operation and management of physical ecosystems. They defined physical ecosystem as a collection of entities and processes that consume physical resources and fulfill functionality. They proposed a data mining framework which uses sustainability metrics such as energy, power consumption and carbon footprint to quantify the desirability of operating the system. Their aim was to monitor the data patterns and determine if it is possible to transform the current operational state to another more efficient one. They mentioned that, domain knowledge required to move from one state to another, where feasible, can be encoded into actionable rules which can be generated through data mining. Several other similar conclusions can be found in (Bash et al., 2008, Marwah et al., 2009b).

In recent years defect diagnosis has become an important activity for ensuring high quality production. It involves the identification of the defect information, which can then be used to adjust manufacturing process accordingly in order to improve the manufacturing yield. Performing an effective defect diagnosis can result in low-cost, high-quality products and improves many other manufacturing performance metrics. DM approach can be very useful here as it is one of the best approaches for analyzing vast amount of operational data. DM defined as the nontrivial extraction of implicit, previously unknown and potentially useful information from data (Frawley, Piatetsky-Shapiro, & Matheeus, 1992) is gaining great interest in economic, manufacturing and scientific domains. DM encompasses a number of different technical approaches such as clustering, data summarization, learning classification rules, finding dependency networks and detecting anomalies (Baykasoğlu & Özbakır, 2007). These DM techniques are used to reveal critical information hidden in the data sets. In this study “classification rule extraction” type DM approach is utilized. Classification rule extraction is a common task of data mining which is characterized by a concern for finding highly predictive rules, often by using heuristic techniques. Classification is the process of finding a set of models or functions which describe and distinguish data classes or concepts, for the purpose of being able to use the model to predict the class of objects whose class label is unknown (Han & Kamber, 2001). Classification formulates a classification model based on the analysis of a set of training data. In classification, a rule generally represents discovered knowledge in the form of IF–THEN rules. Here the main goal of a rule extraction system is to obtain insights for data sets.

In order to carry out an effective data mining study we need an effective data mining algorithm which is able to produce high quality, accurate classification rules. For that purpose, in recent years, there have been numerous attempts for applying several algorithms in data mining to accomplish classification task accurately. Artificial neural network (ANN) is one of these models which are most widely used for that purpose. ANN is a mathematical or computational model based on biological neural networks. It consists of interconnected group of artificial neurons. ANNs are excellent at predicting, learning from experiences, and generalizing from previous examples (Fu & Wang, 2001). Due to this fact, many researches focused on applying ANN in the area of classification rule extraction. However, a main disadvantage of ANN is the difficulty of explaining how and what exactly an ANN has learned. In order to overcome this difficulty, researchers tend to develop new approaches for rule extraction from trained ANNs. Elalfi, Haque, and Elalami (2004) presented a new algorithm for extracting accurate and comprehensible classification rules from databases via trained ANN by using genetic algorithms (GA). Their algorithm is not depended on the ANN training algorithms and does not modify the training results. The GA was used to find the optimal values of input attributes, which maximize the output function of output nodes. They decoded the optimal chromosome and used to get a rule belongs to a target class. Markowska-Kaczmar and Wnuk-Lipinski (2004) presented a method for rule extraction from a ANN based on a GA approach with Pareto optimization. They described the idea of Pareto optimization and shown the details of the proposed method. They tested their method with well known benchmark data sets. Li and Wang (2004) proposed a hybrid system to extract classification rules from decision tables. Differently from the previous studies, in their system ANNs are served only as a tool to reduce the decision table and filter its noises while the final rule set is generated from the reduced decision table by using rough sets. They verified the effectiveness of their approach through experiments and making comparisons with traditional rough set and ANN approaches. Tokinaga, Lu, and Ikeda (2005) studied on the use of ANN rule extraction techniques based on genetic programming (GP) to build intelligent and explanatory evolution systems. They utilized the GP to automate the rule extraction process in the trained ANN where the statements changed into a binary classification. As applications, they generated rules to prediction of bankruptcy and creditworthiness for binary classifications, and applied their method to multi-level classification of corporate bonds by using the financial indicators. Hruschka and Ebecken (2006) proposed a clustering-based approach for extracting rules from multi-layer perceptions type ANN in classification problems. Their rule extraction algorithm basically consists of two steps. First, a clustering GA is applied to find clusters of hidden unit activation values. Then classification rules describing these clusters, in relation to the inputs, are generated. They experimentally evaluated the performance of their approach in four datasets that are benchmarks for DM applications. Setiono, Baesens, and Mues (2009) addressed the generation of comprehensible rule sets from trained ANN. Their algorithm is particularly appropriate in applications where comprehensibility as well as accuracy is required. Their experimental results show that their algorithm produces accurate rule sets that are concise and comprehensible. Özbakır, Baykasoğlu, Kulluk, and Yapıcı (2009) presented a study on rule extraction from trained ANNs for classification problems. The proposed method uses ant colony optimization algorithm for extracting accurate and comprehensible rules from databases via trained ANNs. They experimentally evaluated their algorithm on five benchmark data sets and their results show that the proposed algorithm has a potential to generate accurate and concise rules. Recently authors (Özbakır, Baykasoğlu, & Kulluk, 2010) developed a novel algorithm which is named as DIFACONN-miner for classification rule extraction from ANNs. Differently from the previous approaches DIFACONN-miner integrates ANN-training and rule extraction phases. Therefore, classification rules can be directly obtained from ANNs without needing an extra step for rule extraction. This means that at every ANN-training step the corresponding classification rules are simultaneously generated and training is tried to be achieved for generating more accurate classification rules. In fact, DIFACONN-miner can be called as “rule generation algorithm” instead of “rule extraction algorithm”. DIFACONN-miner uses differential evolution (DE) algorithm for training ANNs and touring ant colony optimization (TACO) algorithm for generating classification rules. Fitness of ANN structure is evaluated according to a multiple objective function which consists of three performance measures namely “error of ANN”, “number of rules” and “training accuracy”. The performance DIFACONN-miner was evaluated by the authors on many different test problems and it was proven that DIFACONN-miner is able to produce accurate and effective classification rules. Therefore we selected DIFACONN-miner as the data miner in the present case study. In the following sections of this paper, first a brief overview of DIFACONN-miner is given. Afterwards the case study and obtained results are presented along with some conclusions.