Detecting hazardous events from sequential data with multilayer architectures

Multivariate time series data play an important role in many domains, including real-time monitoring systems. In this paper, we focus on multilayer neural architectures that are capable of learning high level representations from raw data. This includes our previous solution based on Recurrent Neural Networks with Long Short-Term Memory (LSTM) cells. We build upon this work and present improved methods that aim to achieve higher prediction quality and better generalization to other similar tasks. We apply new deep neural architectures, minimize feature engineering and explore different ways of model selection. In particular, our focus on architectures includes networks with attention mechanism and convolutional networks. We tackle overfitting challenges in a presence of concept drift.