Neural Sequence Analysis Toolbox

This is the latest version of this item.

Abstract

Time series have always been of great interest in the financial sector but today with the advent of sensors and the IoT they have received new attention and their analysis is no longer carried out using linear methods of classical statistics but deep learning is revealing a new paradigm with interesting performances for tasks such as predicting time sequences over time or looking for anomalous patterns that could represent failure of the industrial apparatus. Strategies for time series preprocessing with splines and wavelets are investigated with the present work. Methods such as error based methods and GANs for anomaly detection are also studied and models such as sequence to sequence learning and attention mechanisms for forecasting are taken into consideration. Experiments have been carried out to compare all these methodologies using public data from NASA and airpollution dataset (you can find the links in the experiments chapter). Regarding anomaly detection, the most promising approach was that of GANs. The problem of finding a number of timestamps on which to obtain reliable predictions was also investigated and the problem was formulated in such a way that the neural network itself in the training process can learn the length of the time horizon on which to make predictions. A toolbox has been produced that allows the user to preprocess multivariate time series and implement outlier detection or forecasting applications with the above methodologies.

Available Versions of this Item

• Neural Sequence Analysis Toolbox. (deposited 26 Jan 2023 10:43) [Currently Displayed]