A NOVEL WEIGHTED ELM ENSEMBLE FRAMEWORK FOR CONCEPT DRIFT PREDICTION, DETECTION AND ADAPTION IN UNRELIABLE IOT DATA STREAM

Abstract

Real-time IoT applications that use sensors, and those sensors generate continuous and unreliable data streams.  This IoT stream suffers from Concept drift, which is defined as changes in data distribution over time. Traditional machine learning models fail to adapt to evolving environments, which degrades the model's performance. This paper presents a novel Weighted Ensemble Extreme Learning Machine (ELM) framework that integrates drift prediction, detection, and adaptation into a single architecture. The proposed approach employs an Adaptive Windowing (ADWIN) technique combined with non-parametric statistical tests, where the Kolmogorov-Smirnov (KS) test demonstrates superior performance in drift detection. To handle adaptation, an Online Sequential ELM ensemble is dynamically updated through exponential moving average-based weighting and pruning, ensuring robust generalization and reduced computational overhead. The framework is validated on benchmark datasets, including SEA, Stagger, and a farm weather dataset with artificially induced sudden, gradual, and recurring drifts using Gaussian noise. Experimental results reveal that the proposed method consistently outperforms single classifiers, achieving higher accuracy, precision, recall, and F1-scores, with an average accuracy improvement of up to 97.35% when drift handling is incorporated. These results demonstrate the effectiveness of combining ensemble learning with adaptive drift management. The contributions of this work lie in developing a comprehensive and lightweight framework that unifies prediction, detection, and adaptation of concept drift, thereby enhancing the reliability of IoT-driven decision-making systems. This research provides a scalable foundation for real-time applications such as weather forecasting and can be extended to other domains where streaming data is subject to dynamic changes.