Synthetic time-series data generation using Generative Adversarial Networks
Publicerad
Författare
Typ
Examensarbete för masterexamen
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Data, in this new hybrid era, is the driving force for many industries. The automotive
industry is no exception to this, and the industry has seen an increasing reliance
on data for its operations. The automotive sector, nowadays, offers more services
than just selling a vehicle. It now provides complete mobility solutions like connected,
shared, and autonomous electric vehicles with personalized options. Such
services are possible only with the use of high quality data. Manual data collection
from automobiles is an expensive and laborious task, due to which only a sparse
amount of high quality data is collected. Reduced data means that the operations
and analysis performed are also limited, which makes the company able to offer
less services to the customers. One solution to increase the data is to generate it
synthetically using deep neural networks.
Though there are many methods to generate data synthetically, most of them have
limitations on developing diverse data and preserving the temporal dynamics of the
original data. This thesis focuses on those issues and studies the possibility of designing
a neural network model to generate varied time-series data which has the
characteristics of the original data. In this thesis, a Generative Adversarial Network
(GAN) is implemented to synthetically generate time-series data.
A conventional method of building a machine learning model from scratch is followed
in this thesis, after weighing several factors. The relevant training data is collected
from the vehicle and then pre-processed to improve the quality of the data. Following
this, an initial GAN model is developed that contains the generator and
discriminator structure. Then, an enhanced model with supervised learning mechanism
called timeGAN model is developed for achieving more realistic synthetic
data. This model is then evaluated with suitable metrics, both in a qualitative and
quantitative manner. The thesis aims to resolve the issue of scarce data and thus,
paves the way for effective predictive maintenance of vehicles and better services to
the customers.
Beskrivning
Ämne/nyckelord
Time-series, GAN, Deep learning, Generator, Discriminator, Temporal dynamics, TimeGAN, Data distribution