An Adaptive Hybrid Deep–Statistical Framework for Robust Anomaly Detection in Sequential Systems

Abstract

Anomalies are still difficult to detect in sequential systems due to nonlinearity, temporal dependencies, non-stationarities, noise, and strong class imbalance. This paper introduces an adaptive deep–statistical hybrid architecture for efficient and effective anomaly scoring in diverse sequential setups. It proposes an integration of a sequence autoencoder with BiLSTM architecture with statistical anomaly-detection techniques, such as CUSUM (cumulative sum), EWMA (exponentially weighted moving average), and adaptive thresholding. The autoencoder is used to generate the anomalies based on reconstruction, while the statistical layer analyzes changes in the reconstruction error dynamics. A novel contribution of the architecture is a robustness-aware combination of the deep anomaly evidence from the reconstruction and statistical change evidence in the context of temporal stability. The solution is embodied in a scalable pipeline with support for preprocessing, generation of sequences from time series, anomaly scoring, experiment monitoring, and validation outputs. Testing on four different datasets from the domains of cloud monitoring, financial transactions, cyber security, and business metrics shows that the framework is effective for use cases across industries and is capable of anomaly scoring.