Research: Unsupervised Anomaly Detection in High-Dimensional Multivariate Sensor Stream...

Research question: How effectively can deep generative models detect anomalies in high-dimensional sensor streams compared to traditional statistical approaches?

Modern industrial and IoT environments generate vast streams of high-dimensional sensor data, where early detection of anomalies is crucial for safety, maintenance, and operational efficiency. Traditional anomaly detection methods often struggle with the curse of dimensionality and complex dependencies among variables in such settings.

Recent advances in deep generative models, such as variational autoencoders (VAEs) and normalizing flows, offer new possibilities for modeling the joint distribution of multivariate sensor streams, yet their effectiveness and limitations for anomaly detection in real-world high-dimensional settings remain underexplored.

This research will systematically compare deep generative models to classical techniques like Principal Component Analysis (PCA) and Gaussian Mixture Models (GMM) for unsupervised anomaly detection, using benchmark datasets such as the Secure Water Treatment (SWaT) and NASA Turbofan Engine datasets. The study will evaluate detection accuracy, false positive rate, scalability, and interpretability, and will include ablation studies to understand the impact of dimensionality and sensor correlation.

By rigorously evaluating these methods, the project aims to provide actionable insights for practitioners deploying anomaly detection in high-stakes sensor environments, and contribute to the body of knowledge on scalable, reliable machine learning for real-time monitoring.