A Context-Aware Sensor-Cloud Architecture with Integrated Machine Learning Analytics and Dynamic Model-Orchestration Framework

Abstract

Sensor-cloud monitoring deployments in smart buildings and industrial infrastructure typically evaluate multivariate data streams with fixed, context-independent detection rules. This design cannot detect contextual anomalies — readings that remain within absolute bounds yet are inconsistent with the current operational context — a fault class missed by threshold methods and only partially addressed by context-agnostic machine learning models. We report the design, implementation, and experimental validation of a five-layer Context-Aware Sensor-Cloud Architecture with dynamic Model Orchestration (CASC-MO), in which a dedicated context characterisation layer continuously encodes operational state into a structured context descriptor, and a model-orchestration engine uses this descriptor to select, switch, and adapt anomaly detection models from a per-context registry in real time. The architecture is evaluated on a thirty-day, five-channel multivariate sensor dataset comprising 4,320 observations with four injected anomaly categories: point, contextual, collective, and drift. Compared with a static threshold baseline and a context-agnostic Isolation Forest, CASC-MO achieves an area under the receiver operating characteristic curve (AUC-ROC) of 0.882 versus 0.719 and 0.710, and an overall recall of 82.8% versus 43.8% for both baselines. CASC-MO is the only evaluated approach that detects contextual anomalies, achieving an F1-score of 0.252 in this category against 0.000 for both baselines, directly supporting the core architectural argument. With nearest-neighbour orchestration and a per-context model registry, average inference latency is 17.2 ms per sample, well within the ten-minute monitoring interval used in building automation deployments. By reducing energy waste from undetected heating and ventilation faults and eliminating unproductive maintenance triggered by false alarms, the architecture also supports energy-efficient, climate-responsible building operation. The full implementation and dataset generation pipeline are released to support reproducible research in intelligent sensor-cloud monitoring.