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

Abstract

Internet of Things (IoT) sensor-cloud systems require reliable anomaly detection methods that account for changing environmental and operational conditions. Conventional threshold-based detectors and context-independent machine learning models may fail to identify anomalies whose significance depends on the operating context. This study presents a minimax-based extension of the Context-Aware Sensor-Cloud Architecture with Integrated Machine Learning Analytics and Dynamic Model Orchestration (CASC-MO). The architecture integrates context-specific Isolation Forest models, categorical context descriptors, contamination calibration, and a global fallback model. The proposed minimax selection strategy aims to select a candidate model by minimizing its maximum validation loss across the evaluated operational contexts. The framework is investigated using a simulated smart-building dataset containing 4,320 observations from five sensor channels collected over 30 days. The dataset includes point, contextual, collective, and sensor-drift anomalies. Experimental comparisons are designed to evaluate minimax-based selection against the original Hamming-distance-based CASC-MO method, a static-threshold detector, and a context independent Isolation Forest. Evaluation metrics include precision, recall, F1-score, false alarm rate, area under the receiver operating characteristic curve (AUC-ROC), and inference latency. The study also examines performance across anomaly categories and the limitations of using simulated data. The results are intended to establish whether minimax-based orchestration improves the reliability of anomaly detection across different operating contexts while maintaining acceptable computational cost.