Machine learning for reliability assessment and failure diagnostics in industrial systems with limited data

Abstract

A major objective in prognostics and health management is to understand the reliability of systems throughout their lifespan, with data-driven techniques evolved from classic statistical models to machine learning methods. While data is key to enabling sophisticated methods, practical constraints and limitations in data across industries have not been widely addressed. Therefore, we develop intelligent algorithms for reliability analysis and prognostics by leveraging data with various limitations. First, we introduce a deep learning approach for fleet-level time-to-event (TTE) analysis given limited descriptive covariates. Second, we develop an interpretable approach for subject-specific TTE analysis with explanatory variables while investigating hidden competing risks. Third, we address remaining useful life prediction, handling data scarcity in time-series condition data. Lastly, we further leverage the sparse condition data to detect group anomalies within systems. Our methods achieve competitive performance evaluated with real-world datasets and public benchmark, offering solid insights for reliability analysis and prognostics.