Root Cause Analysis of Product Defects in Manufacturing Using Ensembled Bayesian Networks

Abstract

Root Cause Analysis (RCA) of product defects is crucial to improving manufacturing quality and productivity. Nowadays, manufacturers tend to rely on on-site expert knowledge to identify the root cause of product failure. However, manual RCA is extremely difficult and cumbersome, especially in big data environments yielded by the advancement of information technology and sensor technology. While different model-based methods have been introduced in the literature to localise root causes in a data-driven and automated manner, most of them are prone to various limitations in the aspect of robustness, causality discovery, knowledge representation, stochasticity, and sample size. Therefore, we proposed a product-wise framework of the ensembled Bayesian Network (BN) approach to provide a robust, intelligent and human-interpretable probabilistic reasoning method for RCA to circumvent the issues in the existing techniques. BN is adopted to enable interpretable probabilistic reasoning under uncertainty, which provides reliable decision support for RCA in industrial practice. We developed various structure learning algorithms, a parameter learning algorithm and a Bayesian inference algorithm for BN to learn the root causes of product quality issues from historical product defect records. The Ensemble Learning (EL) techniques enhance BN base learners with bootstrapped re-sampling and combine the predictions from multiple structure learning algorithms, ensuring a robust performance of BN. The structure of the framework is modularised by different products to reduce the sample size and to realise high efficiency. As a result, the proposed method can uncover the causal relationship in the industrial data to support manufacturers to make data-driven decisions under the circumstances of product quality failures. To achieve such goals, this project has automatically acquired causal knowledge, identified the root cause with probabilities and predicted quality risks in production. The proposed method has been implemented on real-world data collected from the plastic industry. Experimental results have shown that the ensembled BN framework successfully discovers the root cause along with corresponding probabilities and predicts the poor-quality instances with considerable robustness and accuracy.