Master's Thesis : State segmentation and forecasting of production processes by machine learning

Abstract

In this thesis, two different tasks concerned with time series of production processes are tackled. The first one is a time series segmentation into different classes for each timestep of the time series. An accurate timestep classification has multiple useful applications such as doing retrospective analyses of production processes based on the percentage of occurrence of each class or implementing an intelligent system to activate or deactivate alarms applied to processes based on their current state. Then, the second task is a multistep multivariate time series forecasting*. It can be used to forecast specific events in order to avoid them or to prepare for them.

This work focuses on the application of machine learning algorithms to those two problems with the objective to automate and generalize the solution to the broadest range of production datasets as possible. The end goal is to study the potential of those algorithms, rather than delivering a perfect solution.

For the time series segmentation, tree-based models are considered. In the final evaluation, they display an irregular performance alternating between very high and low accuracy depending on the classes. However, the lack of precision might be caused by an external bias in the data labelling. Still, its performance on the best classes reveals its high potential.

For the time series forecasting, the study focuses on deep learning algorithms which gave good results in this domain. Two state of the art level models are tested: DeepAR and Temporal Fusion Transformers (TFT). The evaluation demonstrated the difficulty encountered by the models and, by extension, the difficulty of an automated timestep forecasting of a wide range of datasets by deep learning.

*Multistep time series forecasting denotes the fact of predicting multiple timesteps of a time series while the multivariate term indicates that forecasts are done on more than one value for each timestep.