Abstract
Abstract
Specific industry-level chemical reactions taking place
> inside high-pressure containers give rise to complex nonlinear systems
> that can occasionally get out of control leading to significant damage
> of assets of tremendous economic value. Utilizing sophisticated
> machine learning models based on temporal readings of physical
> quantities through sensors within these chambers, we intend to predict
> the throughput to be obtained out of the reaction and to detect
> anomalous behavior of the system, indicating a compelling probability
> of runaway in real-time. Analogous predicaments are encountered in
> numerous industries where production of goods is monitored via
> multiple sensors, giving rise to vast collections of multivariate
> time-series data. Even though being a prominent challenging topic for
> several decades within the machine learning community, time-series
> mining recently gained more interest in the light of blossom in deep
> learning domains where effective automated solutions to issues
> involving huge amounts of data exhibiting sequential dependencies. As
> public repositories of multivariate time-series data sets, including
> sensory data, are established, neural network models based on
> variations of Convolutional and Recurrent Neural Networks have been
> emerging and claiming to attain state-of-the-art performances in a
> multitude of problems. However, concerning the time-series field, the
> accuracy gap between deep learning methods and traditional machine
> learning approaches is not as significant as in the computer vision
> area, suggesting a further increase in research activities and room
> for development in the foreseeable future. Our study aims to take part
> in the impending progress through devising effective and flexible
> prototypes.