Domain-adapted explainability for machine learning predictions of rotodynamic pump degradation in safety-critical industrial sectors

In safety-critical industries, it is essential to have clear and trustworthy predictive models to ensure reliability and build confidence. This paper presents a new framework to explain predictions made by Machine Learning (ML) and Artificial Intelligence (AI) models, specifically designed for experts who may not have technical knowledge of these technologies. The focus is on predicting potential issues in pumps that play a critical role in moving fluids within industrial systems. The framework uses real-world data and a tool called Shapley Additive exPlanations (SHAP) to explain how different factors influence the model’s predictions. These explanations are transformed into clear, easy-to-understand text and visuals, making them accessible to users without technical expertise. The framework was tested on predicting pump performance issues and demonstrated its ability to build trust by aligning explanations with existing expert knowledge. By offering accurate and reliable insights, this approach supports the adoption of ML tools in industries with strict regulations, fostering confidence in their use for critical decision-making.