Integrated Unsupervised Anomaly Detection in Large-Scale Hospital Electrical Monitoring Data Using Isolation Forest and Autoencoder

The increasing volume and dimensionality of electrical monitoring data in healthcare facilities create challenges for identifying unusual operating behavior without verified fault labels. This study presents a reproducible unsupervised framework for large-scale hospital electrical monitoring data combining data-quality screening, Isolation Forest (IF), Autoencoder (AE), threshold-sensitivity analysis, and event-level engineering evaluation. The original dataset contained 3,429,555 observations and 103 electrical variables; after quality screening and removal of one constant variable, 3,076,297 observations and 102 variables were retained. At the main operating threshold, IF identified 29,045 observations (0.9442%) and AE identified 34,213 (1.1121%), including 5759 jointly identified observations. The Jaccard similarity decreased from 0.1343 at the strict threshold to 0.1002 at the main threshold and 0.0836 at the broad threshold, indicating greater agreement among the most extreme observations. At the main threshold, 57,499 union observations formed 14,646 temporally contiguous events. Event-level analysis revealed deviations in power, current, voltage, power factor, and harmonic-distortion measurements, while approximately 88–91% of evaluated event-variable deviations moved closer to their local pre-event baseline afterward. Although verified fault labels were unavailable, these results provide within-dataset engineering evidence that many model-identified events corresponded to temporally coherent and electrically observable changes.