Master's Thesis : Time series analysis with machine learning

Abstract

Many industrial companies have production chains that run in batches, as opposed to a continuous production chain. The monitoring of their production chains is crucial and yields time series data. PEPITe, a data analytics company, works which one such industrial company in particular whose yield in one batch production chain fluctuates without apparent reason. So far, PEPITe’s arsenal of traditional machine learning and statistical methods has not provided sufficient results. This work’s objective consists of two parts. The first part aims at finding unsupervised algorithms on time series data that PEPITe could add to their tool box. The second part is more specific to the problem of the industrial company, which aims at finding a supervised method capable of providing interpretable results. Regarding the first objective, 4 algorithms were selected from the literature, implemented and showcased on the provided industrial data set. Regarding the second objective, two interpretable methods, namely a convolutional neural network (CNN) and a classifier based on shapelets, were tested on the provided data set and compared against 2 baselines. The implementation of the unsupervised algorithms features code with permissive licenses to be able to build upon this work. The performance of all the supervised models was lower than expected, however the CNN allowed for a kind of interpretability with a comparable performance to the baselines.