An Interpretable Fault Prediction Method Based on Machine Learning and Knowledge Graphs

Predictive Maintenance (PdM), rooted in data analysis and advanced technology, is a maintenance strategy that predicts the likelihood of future failures by monitoring the operational status of equipment or systems. It aids engineers in proactively formulating maintenance strategies. However, its opaque decision-making process poses challenges in providing comprehensible and rational explanations, leading to user distrust in predictive outcomes. To address this issue, we propose an interpretable fault prediction approach based on Machine Learning (ML) and Knowledge Graphs (KG). This approach not only diagnoses early equipment faults and identifies anomalous components but also furnishes explanations. Initially, appropriate fault prediction models are selected through experimentation to ensure the accurate diagnosis of equipment status. Subsequently, to attain comprehensive and accurate explanations, we amalgamate two interpretable methods, Shapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), to explicate the predictive results. Finally, an equipment KG is constructed based on experimental datasets to pinpoint faulty components and further elucidate predictions.