Improving nonconformity responsibility decisions: a semi-automated model based on CRISP-DM

Abstract

Nonconformity (NC) management is a fundamental process in production, yet the literature notion of it does not always align with what is practiced in reality. In particular, the literature often excludes the NC responsibility decision, which is a difficult, costly and time-consuming task assignment, but also an integral part of the NC management process. We propose a semi-automated model we call SANC, which improves the accuracy of NC responsibility decisions and significantly cuts their costs. We base our methodology on CRISP-DM and extend it to fit the semi-automated NC responsibility decision. Unlike the original CRISP-DM, SANC utilizes existing organizational resources, and thus extends the capabilities of CRISP-DM in terms of both achieving greater overall performance and broadening its appeal to more traditional production processes. We demonstrate this solution by implementing it in a large-scale assembly plant in the printing industry, that may result in savings of over $186 K according to our assessments.

Appendix

• 1. ML models overview

  • 1.1 Decision trees

    Simple classifiers that consist of a collection of decision nodes arranged in a tree structure. It is an “old-timer” with successful implementations in chemical and biological problems (Bjij et al. 2016), network component failure analyses (Winkler et al. 2018) and many more (Wu et al. 2016). We tuned a single hyper-parameter: the depth of the tree.

  • 1.2 Support Vector Machines (SVM)

    A group of classification methods that has been successfully implemented in image processing classification (Noi and Kappas 2018), production control (Gao and Hou 2016) and others. We tuned the following hyper-parameters, based on Noi and Kappas’s study (Noi and Kappas 2018): the kernel, C—cost of misclassification, and γ—width.

  • 1.3 Random Forest (RF)

    A methodology that relies on a combination of many decision trees. RF has been successfully implemented in various fields (Alipour et al. 2017; Mokrova et al. 2018). The hyper-parameters tuned, following Scornet’s study (Scornet 2017), were the number of features in a tree, number of trees and minimum samples in a leaf.

  • 1.4 Artificial Neural Networks (ANN)

    A method that originally attempted to mimic human brain information processing (Hosseini and Khaled 2019). ANN has been successfully applied in many fields for various classification and forecasting problems (Kan 2017; Kwon et al. 2019). Hyper-parameters tuned based on heuristics, introduced in Sheela and Deepa’s review (Sheela and Deepa 2014), were the number of hidden layers, number of neurons in each layer and the activation function.

  • 1.5 Multinomial Regression (MLR)

    An extension of binary logistic regression that allows classification of more than two categories (Starkweather and Moske 2011). MLR is often perceived as an attractive analysis because it requires no assumptions to be made. Variations of it have been applied in a wide range of subjects, including image processing and various medical predictions (Hogland et al. 2017; Mazzocco and Hussain 2012). The ridge parameter lambda was hyper-tuned.