Explainable AI-Based Predictive Maintenance for Operator Decision Support

Sammanfattning

Introduction: Industrial production relies on complex machine-intensive systems where equipment failures can cause significant operational and economic consequences. Predictive maintenance (PdM) aims to anticipate failures using data-driven methods, but many models lack transparency, making their outputs difficult for operators to interpret and trust. Explainable AI (XAI) has been proposed to address this issue, yet empirical research on its role in operator decision-making remains limited. Research Question: How can explainable AI-based predictive maintenance support operator decision-making by improving understanding of machine behavior in industrial production? Method: This study adopts a qualitative case study approach in collaboration with an industrial company. Data were collected through two semistructured interviews with engineers responsible for machine monitoring and maintenance. The material was analyzed using structured qualitative content analysis, combining deductive coding with inductive pattern identification. Results: The findings suggest that the interviewed engineers primarily relied on experiential knowledge when interpreting machine behavior and making maintenance decisions. The findings further suggest that predictive insights were perceived as most useful when they provided clear indications of abnormal behavior, timing of failures, and relevant influencing factors. Explainability was described as supporting the interpretation and evaluation of predictions rather than replacing human decision-making through automation. Discussion: The findings suggest that explainable predictive maintenance can support decision-making by translating model outputs into understandable and context-relevant information. Rather than replacing human expertise, explainable AI functions as decision support that enhances understanding and reduces uncertainty. However, the study is limited by its small sample and single-case design. Future research could examine different explanation approaches across multiple industrial contexts and their impact on trust and decision quality