AI-Driven Anomaly Detection in IoT Time Series: A Hybrid Approach to Classification and Feature Extraction

Abstract

Anomaly detection in Internet of Things (IoT) systems plays a crucial role in identifying failures, security breaches, and abnormal operational patterns across a wide range of applications, from smart cities to industrial automation. Given the high volume, velocity, and variability of IoT time series data, traditional detection models fall short in terms of scalability and accuracy. This paper proposes a hybrid AI-based framework that combines advanced feature extraction techniques with multi-stage classification models for real-time and context-aware anomaly detection in IoT time series data. The methodology integrates statistical feature engineering, deep autoencoders, and hybrid classifiers—including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and ensemble learning methods—to capture temporal dependencies and contextual variations in heterogeneous sensor environments. Experimental evaluations across benchmark IoT datasets demonstrate superior detection accuracy, reduced false positives, and enhanced robustness in noisy environments. The results confirm that hybrid models significantly outperform single-algorithm approaches in detecting subtle and complex anomalies, making them well-suited for dynamic IoT ecosystems.