Machine learning has become a part of our daily life and is commonly used across a wide range of industries, these methodologies have been applied in countless areas of application and their use is in continuous expansion. In particular, these approaches play a key role in enabling Industry 4.0 and IoT scenarios. Many of the algorithm results cannot be understood and explained in terms of how and why a specific decision was made. With the advancement of machine learning research, several techniques and approaches have emerged in recent years, but only a few studies have been produced regarding the end-user perspective and understanding the results of the algorithms, making these algorithms into higher beings in which their credibility depends on how much faith one has. Therefore, the lack of interpretability in this technology is the biggest obstacle to the spread of these applications. Anomaly detection is a large subdivision of machine learning technology that has enormous applicability in industrial scenarios. In fact, it is extremely relevant for the purposes of quality monitoring, predictive prevention and much more. Furthermore, the strength of this type of approach is that it can be implemented without the need for tagged data and obviously in this type of frameworks where the data is "dirty", is very peculiar not to have labeled data. Obviously, this last application is also infected with the same problem that the whole family suffers from. This thesis describes the development of an anomaly detection system that is interpretable, which therefore aims at alleviate the problems introduced above by trying to focus as much as possible on the perspective of the end-user. The two main topics are anomaly detection on the one side and the interpretability of the models on the other.
Machine learning has become a part of our daily life and is commonly used across a wide range of industries, these methodologies have been applied in countless areas of application and their use is in continuous expansion. In particular, these approaches play a key role in enabling Industry 4.0 and IoT scenarios. Many of the algorithm results cannot be understood and explained in terms of how and why a specific decision was made. With the advancement of machine learning research, several techniques and approaches have emerged in recent years, but only a few studies have been produced regarding the end-user perspective and understanding the results of the algorithms, making these algorithms into higher beings in which their credibility depends on how much faith one has. Therefore, the lack of interpretability in this technology is the biggest obstacle to the spread of these applications. Anomaly detection is a large subdivision of machine learning technology that has enormous applicability in industrial scenarios. In fact, it is extremely relevant for the purposes of quality monitoring, predictive prevention and much more. Furthermore, the strength of this type of approach is that it can be implemented without the need for tagged data and obviously in this type of frameworks where the data is "dirty", is very peculiar not to have labeled data. Obviously, this last application is also infected with the same problem that the whole family suffers from. This thesis describes the development of an anomaly detection system that is interpretable, which therefore aims at alleviate the problems introduced above by trying to focus as much as possible on the perspective of the end-user. The two main topics are anomaly detection on the one side and the interpretability of the models on the other.
Machine Learning approaches for Anomaly Detection in Industrial IoT scenarios
CONVENTO, ENRICO
2021/2022
Abstract
Machine learning has become a part of our daily life and is commonly used across a wide range of industries, these methodologies have been applied in countless areas of application and their use is in continuous expansion. In particular, these approaches play a key role in enabling Industry 4.0 and IoT scenarios. Many of the algorithm results cannot be understood and explained in terms of how and why a specific decision was made. With the advancement of machine learning research, several techniques and approaches have emerged in recent years, but only a few studies have been produced regarding the end-user perspective and understanding the results of the algorithms, making these algorithms into higher beings in which their credibility depends on how much faith one has. Therefore, the lack of interpretability in this technology is the biggest obstacle to the spread of these applications. Anomaly detection is a large subdivision of machine learning technology that has enormous applicability in industrial scenarios. In fact, it is extremely relevant for the purposes of quality monitoring, predictive prevention and much more. Furthermore, the strength of this type of approach is that it can be implemented without the need for tagged data and obviously in this type of frameworks where the data is "dirty", is very peculiar not to have labeled data. Obviously, this last application is also infected with the same problem that the whole family suffers from. This thesis describes the development of an anomaly detection system that is interpretable, which therefore aims at alleviate the problems introduced above by trying to focus as much as possible on the perspective of the end-user. The two main topics are anomaly detection on the one side and the interpretability of the models on the other.File | Dimensione | Formato | |
---|---|---|---|
Convento_Enrico.pdf
accesso aperto
Dimensione
6.44 MB
Formato
Adobe PDF
|
6.44 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/35225