@mastersthesis {pub4109,
	title = {Real-world Anomaly Detection by using
Digital Twin systems and Deep Learning
algorithms},
	author = {Andrea Castellani},
	year = {2019},
	abstract = {Nowdays, with the continuously growing of monitored data present in Smart Company environment, the
need of some anomaly detection technique is became more relevant, in order identify some anomalous behavior
from time series data generated by sensors. With a Digital Twin, a detailed simulation of a complex physical
object or system, is possible to provide more data to feed in a machine learning algorithm and thus help in
the anomaly detection task. This report presents the work done during six months internship at the Honda
Research Institute Europe; the goal of the project was to use the Digital Twin of some machines, to detect
their anomalous working operation in the real world. This paper explains the Smart Company environment
present in the building and its monitoring system and then illustrate the current Digital Twin in such a complex
environment. So, it shows an unsupervised approach to anomaly detection using simulation data from the
Digital Twin to train some articial neural network (Autoencoders), with the aim of detecting anomalies in
a real-world data set. The data are gathered both from Smart Meters present in the Company and they
are simulated by the Digital Twin. Some statistical and contextual features are extracted from them before
feeding the learning algorithms. So, a novel method, based on Deep Siamese Autoencoder, is presented; it
is aimed to improve the performance of such unsupervised algorithms with the awareness of a rather small
subset of labeled data. Furthermore, an end-to-end machine learning model is described, able to work with
pure raw data. De facto, it has the same structure of the Siamese Autoencoder but uses convolutional neural
networks to extract automatically meaningful features. All of those developed algorithms are compared with
the state-of-the-art approach for anomaly detection and the overall results are very promising. They show
that the presented methods, during testing, reach an ROC AUC and F 2 score noticeably higher the recent
state-of-the-art anomaly detection techniques chosen for comparison.},
	publisher = {Universit{\`a} Politecnica delle Marche Home},
	booktitle = {Univ. delle Marche, Ancona}
}