Operations Research 2025
Abstract Submission

2231. Uncovering Factors Affecting Manufacturing Quality: Causal Discovery from Embedded Features in Manufacturing Data

Invited abstract in session TA-6: Prescriptive Analytics: Sales, Logistics and Industrial Analytics, stream Analytics, Data Science, and Forecasting.

Thursday, 8:45-10:15
Room: H9

Authors (first author is the speaker)

1. Won Sang Lee
Dept. of Data Science, Gangneung Wonju National University

Abstract

Recently, there is growing interest in data-driven quality improvement and defect reduction using AI. Many recent approaches rely heavily on deep learning techniques that use large numbers of features. Unfortunately, these methods often experience limited interpretability and lack efforts to uncover the systematic relationships among features. However, in manufacturing, features are not isolated; they are generated through the interaction of facility configurations, environmental variables, and operational settings. These factors influence outcomes via intricate pathways. Thus, identifying the causal structure among features could offer more actionable and effective insights for process optimization. To address this, a novel approach is proposed to integrate embedding representation with causal discovery. First, embedding techniques allow for the transformation of fragmented feature data into compact, informative representations, since the sensor-derived features are often high-dimensional, fragmented, and difficult to interpret. Once embedded, these embeddings are analyzed using causal discovery to reveal underlying causal pathways and estimate their effect. Then, the proposed method is empirically applied to the dataset concerning failure events in rotating machinery. The sensor-based features are embedded to represent the operational state of machines, and causal inference is conducted on these embeddings. The results effectively identify important embeddings that significantly contribute to failure events and uncover the causal relationships among them. The proposed framework provides a foundation for scalable and interpretable quality improvement strategies in future manufacturing research.

Keywords

Status: accepted

Back to the list of papers