Multi-Feature Classification of Displacement Alarm Causes Using IoT-Based Slope Monitoring Sensor Data

Internet of Things (IoT)-based slope monitoring systems support real-time stability surveillance, but frequent false alarms cause alarm fatigue and delay responses to genuine hazards. We developed a clustering-based framework that classifies alarm causes into three categories—Malfunction, Actual Displacement, and Others—from inclinometer and extensometer data. From 55 alarm events recorded at 43 monitoring stations in 2025, 85 domain-informed features (statistical, waveform, spatial-correlation, trend, sudden-synchronization, and malfunction-pattern) were extracted and compared across five feature-extraction methods, six selection levels, and three clustering algorithms. Because the feature ranking uses expert-assigned labels, the framework is characterized as label-informed feature ranking followed by unsupervised clustering, and it was evaluated with 100 random seeds, nested leave-one-site-out validation, bootstrap confidence intervals, and permutation tests. A compact set of five features—three cross-correlation-based synchronization features and two spatial-correlation features—was stable across seeds and achieved a leave-one-site-out accuracy of 0.80 (95% CI 0.69–0.89) with the full 8 h window. Using only data recorded up to the alarm time, accuracy rose to 0.91 (0.84–0.98), 12 of the 13 Actual Displacement events were recovered (recall 0.92), and no non-displacement event was assigned to the displacement cluster. Synchronization features separated genuine displacement from the other causes, whereas spatial-correlation features distinguished external disturbance from sensor malfunction. The interpretable framework is proposed as a decision-support filter that complements conventional threshold-based alarms.