Contextual isolation forests: contextual anomaly detection in water distribution networks using semantic partitioning

Anomaly detection in water distribution networks (WDNs) using unsupervised methods faces a fundamental limitation: the inability of the classic Isolation Forest (IF) algorithm to locate the semantic origin of detected anomalies and quantify their severity in an actionable manner within environments whose operational nature is constantly evolving. This paper proposes Contextual Isolation Forest (CIF), an adaptive extension of IF designed to operate on dynamic systems with continuous sensor data streams. CIF integrates three sequential procedures: domain-specific formal data modelling, empirical estimation of hyperparameters via controlled contamination, and contextual explainability through semantic partitioning based on domain expert criteria, functional properties of the system and geographical distribution. The method introduces two new anomaly indicators, Recurrence (RI) and Intensity (II), which extend the output of the classic IF from < score > to < score, RI, II>, enabling prioritisation and diagnosis without additional post-hoc analysis, and laying the foundations for self-adaptive monitoring strategies. The model was validated on 33,096 real-world records, derived from 11 sensor variables distributed across 4 geographical zones of a WDN in south-eastern Spain. The results demonstrate the stabilisation of detection metrics with just 10 trees and a threshold of 0.70, identifying 5 critical anomalies (RI ∈ [0.38, 0.63] and II ∈ [0.00, 0.77]) with explicit localisation by zone and type of measurement. CIF complements IF, EIF and DIFFI by adding operational explainability through semantic subspaces, without requiring labels or assuming prior distributions, constituting a viable component for anomaly detection systems in continuously evolving industrial environments.