Explainable and Data-driven AI-Methods for Predictive Maintenance

Sammanfattning

This study presents a data-driven approach for predictive maintenance in complex technical systems by applying survival analysis and machine learning to operational sensor data. The study investigates how these methods can improve time-to-failure prediction and support maintenance-related decision-making. To address the research question, a structured methodology based on the CRISP-DM framework was applied. This study used two public datasets, Component X from SCANIA and CMAPSS from NASA. Survival analysis and Random Survival Forest were used to predict failure and risk over time. Both of these methods were evaluated with the Concordance Index. Explainable AI methods were also included, where SHAP identified the most influential variables and Counterfactual Explanations showed how changes in operational conditions could extend remaining useful life. The results showed that the models identified differences in failure risk between units. SHAP revealed that only a few sensors had a significant predictive impact, while Counterfactual Explanations showed that small changes in operational variables could extend the remaining useful life. The methods were effective in modeling both risk and time to failure. In addition, the study highlights the relationship between risk scores and remaining useful life, reflects the complexity of real-world systems. Results show how machine learning and explainable AI can contribute to both predictive performance and interpretability in maintenance decision support. The use of public datasets may, however, limit the generalizability of the results, and future research should validate the methods using real-world operational data.