Enhancing Time Series Anomaly Detection Using Discretization and Word Embeddings


Time series anomaly detection plays a pivotal role across diverse fields, including cybersecurity, healthcare and industrial monitoring. While Machine Learning and Deep Learning approaches have shown remarkable performance in these problems, finding a balance between simplicity and accuracy remains a persistent challenge. Also, although the potential of NLP methods is heavily expanding, their application in time series analysis is still to be explored, which could benefit greatly due to the properties of latent features. In this paper, we propose WETAD, a novel approach for unsupervised anomaly detection based on the representation of time series data as text, in order to leverage the use of well-established word embeddings. To showcase the performance of the model a series of experiments were conducted on a diverse set of anomaly detection datasets widely used in the literature. Results demonstrate our approach can compete and even outperform state-of-the-art approaches with a simple, yet effective model.

In International Conference on Soft Computing Models in Industrial and Environmental Applications
Nahuel Costa
Nahuel Costa
Machine Learning & AI researcher