nirojasva/spectral-benchmark-bv4

Spectral Benchmark BV4 for Real-Time Anomaly Detection in High-Dimensional Data Streams

Abstract

This paper introduces the Spectral Benchmark BV4, a high-dimensional spectral dataset designed for unsupervised anomaly detection in data streams within vacuum environments. Collected using an Optical Emission Spectroscopy (OES) sensor, the dataset comprises 9 experiments representing practical scenarios. The challenges include spatial anomalies, temporal anomalies, and concept drift. BV4 captures 2048 distinct wavelengths over time. This provides a rigorous framework for evaluating the ability of machine learning models to detect abnormal patterns under dynamic environmental changes in vacuum chamber conditions.

Dataset Structure

The dataset contains 9 CSV files representing different anomaly scenarios.

• SA refers to Spatial Anomalies.
• TA refers to Temporal Anomalies.
• DA refers to Concept Drift.

Responsible AI Metadata

The extended Croissant metadata files containing detailed data limitations, biases and provenance information for each experimental scenario (DA, SA, TA) are provided as independent JSON files in the metadata/ directory of this repository. Researchers should refer to these files for rigorous evaluation constraints.

Academic Citation

You can cite the original work using the following BibTeX format:

@inproceedings{varela2026onlinebootknn,
  title={The BV4 Benchmark for Unsupervised Anomaly Detection in High-Dimensional Spectral Data Streams},
  author={Rojas Varela, Nicolas and Ah-Pine, Julien and Mephu Nguifo, Engelbert},
  booktitle={Advances in Neural Information Processing Systems},
  volume={},
  number={},
  pages={},
  year={2026}
}