Multi-task Bayesian network learning with heterogeneous data mode for root cause analysis in multi-variety small-batch manufacturing

Purpose The purpose of this study is to develop a root cause analysis (RCA) solution for quality failures in complex industrial settings under multi-variety, small-batch production. It aims to systematically address key challenges in data-driven approaches, including data scarcity, heterogeneous cross-variety data distributions, and the coexistence of scalar and functional variables. Design/methodology/approach This study proposes a Heterogeneous-data-based Family Bayesian Network (H-FamilyBN) learning scheme for cross-variety RCA. Specifically, the problem is cast as a multi-task learning formulation over a family of Bayesian networks (BNs), each corresponding to a distinct product variety. A distribution-adaptive data augmentation strategy is first employed to construct task-specific datasets. Then, considering prior domain knowledge, each BN structure is oriented by defining pairwise directed edge learning with scalar and functional variables. Meanwhile, a causality divergence penalty is introduced to facilitate information sharing across family BN members. Findings A case study on robotic welding in automotive production is conducted to verify the validity of the proposed scheme. The scheme demonstrates interpretability and effectiveness by capturing topological structure and revealing the underlying causality among variables. Additionally, it achieves higher prediction accuracy and generates denser BN structures compared with the benchmark methods, thereby decreasing the likelihood of false negatives in causal inference and fault diagnosis. Originality/value This study introduces an intelligent decision-making scheme that integrates industrial data with domain knowledge for quality management. The scheme offers a valuable tool for identifying the faults and their propagation paths by leveraging the BN technique, providing practical insights for process improvement and quality enhancement.