On the reusability of machine learning-based process monitoring systems for manufacturing digital twins

Advanced manufacturing is increasingly supported by machine learning (ML)-based digital twins (DTs) for real-time process monitoring and quality assurance. However, changes in physical domain configurations (e.g., machines, materials, and sensors) often cause domain shifts, limiting the reusability of existing DT components. Rebuilding DTs from scratch for each new configuration is costly and time-consuming. To address this challenge, we define DT reusability through three key criteria: FAIRness (findability, accessibility, interoperability, and reusability), transferability, and transfer efficiency. We propose a framework to systematically support the reuse of ML-based process modeling components in DTs, consisting of three phases: FAIR compliance, transferability analysis, and domain adaptation. To enhance transfer efficiency, we introduce the domain-adversarial and decision distribution alignment (DADDA) network, which enables class-conditional alignment and mitigates overfitting through competing domain alignment objectives. A case study on vision-based process monitoring in additive manufacturing was conducted to validate the proposed framework. A FAIR-compliant database of existing DT components was developed, and the most suitable source domain for the designated target domain was identified through an evaluation of semantic and statistical similarity. Leveraging the selected source dataset, DADDA achieved 84 % accuracy after unsupervised pre-training and 96.9 % after supervised fine-tuning with only 210 labeled examples. Further validation on acoustic-based monitoring systems demonstrated the applicability of DADDA to various modalities. • Changes in physical domain configurations of a digital twin degrade process modeling performance. • Reusability is crucial for maintaining the effectiveness of digital twins over such changes. • Systematic reuse is achieved by ensuring FAIRness, transferability, and transfer efficiency. • Proposed unsupervised domain adaptation improves transfer efficiency while mitigating overfitting. • Validation on laser additive manufacturing shows high accuracy with minimal labeled data required.