Anomaly and Novelty Detection in Non-Stationary Time Series: Deep, Hybrid, and Federated Perspectives

As cyber-physical systems, the Industrial Internet of Things (IoT), and ongoing health monitoring become increasingly embedded in daily operations, we’ve seen a surge in multivariate time series (MVTS) data, underscoring the critical role of anomaly detection (AD) within intelligent system designs. However, today’s real-world applications often face extreme non-stationarity, shifts in means, variances, and covariances over time, driven by factors such as sudden changes, seasonal cycles, sensor wear, or equipment aging. In this chapter, we provide an in-depth review of methods for spotting anomalies and novelties in such unsteady time series, weighing their strengths and weaknesses around flexibility, clarity, and processing demands, with a special focus on edge computing setups. Pulling from both classic and cutting-edge studies, we outline three main paths forward: first, crafting hybrid AD systems that blend Variational Autoencoders (VAE) and Transformer models with Support Vector Data Description (SVDD), Multivariate Exponentially Weighted Moving Average (MEWMA), and Reinforcement Learning (RL) to deliver steady, adaptable performance, much like recent efforts that sharpen reconstruction accuracy and fine-tune decision boundaries; second, developing lightweight versions through quantization, pruning, and FPGA boosts to tackle edge hardware limits, allowing quick hyperspectral anomaly checks with low delays; and third, pushing forward group-based AD in federated learning contexts that handle varied client drifts, promoting secure, spread-out analysis while tackling non-IID issues via grouped setups.