A 3-level interpretable data-driven framework for operational analysis of industrial centrifugal pumps

Abstract This paper proposes an interpretable framework for industrial condition monitoring, enabling a three-level understanding of machinery behaviour (machine, context, and system). Using multivariate time-series from a WirelessHART sensor network, the framework combines transparent data-mining and statistical-learning components (density-based clustering, principal component analysis, variable-importance modelling, and time-series dependency analysis) to extract readable patterns and rules describing operational dynamics. Unlike approaches that treat machine, context, or system levels in isolation, the proposed workflow unifies them within a single analytical scheme built around intrinsically interpretable modules. At the machine level, DBSCAN and PCA identify and characterise distinct operational states and enable compact decision rules; at the context level, the Kolmogorov-Smirnov test and a consensus ranking of variable importance quantify campaign-to-campaign drift and highlight the role of environmental variables; at the system level, Granger test and cross-correlation analysis identify short-lag directional predictive relationships between concurrently operating machines. The framework is demonstrated on three centrifugal pumps monitored across multiple campaigns in an industrial plant. All data are publicly available, enabling full replication.