This project consisted on applying deep learning to time series anomaly detection.
The proposed approach is based on a Deep Generative Model - the Variational Autoencoder - and integrates recurrent neural networks and an attention mechanism.
- UNSUPERVISED - no need for big labelled datasets. Just takes in the raw data and it learns everything from scratch.
- GENERIC - Suitable for every type of time series data (seasonal, non-seasonal, predictable, unpredictable.). Can also be applied to other sequential data structures, such as text and videos.
- SCALABLE - Inference and the computation of the anomaly scores are efficient, both taking in total a few milliseconds.
Conference Paper: Unsupervised Anomaly Detection in Energy Time Series Data using Variational Recurrent Autoencoders with Attention
Oral Presentation in the 17th IEEE International Conference on Machine Learning and Applications (ICMLA'18).
PDF (accepted version)