OnlineBootKNN: A Novel Algorithm for Detecting Anomalies in Spectral Data Streams
Résumé
Monitoring the elemental composition of materials in real-time is essential for many real-world applications, including quality control in manufacturing, environmental monitoring, and space exploration. This continuous monitoring can be achieved by analyzing a flux of spectral data, known as a spectral data stream, which consists of a sequence of instances, each representing a set of wavelengths and their associated intensities. While state-of-theart anomaly detection methods in a streaming setting have primarily focused on tabular data, their applicability to spectral data streams remains underexplored. To address this gap, we evaluate the performance of existing tabular anomaly detection methods, including decision tree-based, deep learning-based, density-based, and distancebased approaches, on spectral data streams. Further, we introduce OnlineBootKNN, a novel unsupervised approach that combines knearest neighbors with online bootstrapping to detect anomalies in spectral data streams, creating a robust framework for identifying anomalous spectral instances in real-time. We demonstrate the high performance of OnlineBootKNN using real-world datasets, benchmarking it against state-of-the-art anomaly detection techniques for tabular data. Additionally, we evaluate the efficiency of our method relative to other leading approaches. Finally, we emphasize the inherent interpretability of our framework, which is especially valuable in scenarios where identifying the most relevant wavelengths in the spectra is crucial for differentiating between normal and anomalous instances.