This thesis examines the use of machine learning and deep learning in mechanical engineering problems. An underlying theme of the research is the comparison of Convolu- tional Neural Networks and Recurrent Neural Networks with standard machine learning techniques, such as Support Vector Machines, showing the advantages of Deep Learning in modeling complex phenomena where traditional approaches fall short. As an example, it is shown that with the proposed Multichannel Spectrograms, an autonomous robot can classify the traversed terrain with an accuracy of over 90% when using a CNN based on proprioceptive signals. This work addresses both generalization and extrapolation issues, demonstrating that Deep Learning delivers more robust results than standard machine learning when applied to data obtained in varied conditions. The importance of sensor data complementarity is also emphasized showing that the use of appearance- based measurements can be used to predict wheel-terrain interactions, improving the ability of planetary exploration rovers to avoid hazards and respond to changing conditions. The combination of proprioceptive and exteroceptive signals is explored for crop monitoring in agricultural robotics. The design and development of a multi-sensor system for 3D reconstruction of agricultural environments is here detailed together with the necessary sensor fusion algorithms. Additionally, a new unsupervised learning algorithm, Kalman Supervised Network, is proposed for modeling mechanical systems. KSN uses the principles of Kalman Filters to continuously learn from sensor measurements and produce a more accurate model of the system than the one embedded in the KF.
Questa tesi esamina l'uso del machine learning e del deep learning nei problemi di ingegneria meccanica. Un filo conduttore della ricerca è il confronto di Convolutional Neural Networks e Recurrent Neural Networks con tecniche di machine learning standard, come le Support Vector Machines, mostrando i vantaggi del Deep Learning nella modellazione di fenomeni complessi dove gli approcci tradizionali falliscono. A titolo di esempio, si dimostra che con gli spettrogrammi multicanale proposti, un robot autonomo può classificare il terreno attraversato con una precisione superiore al 90% quando utilizza una CNN basata su segnali propriocettivi. Questo lavoro affronta sia i problemi di generalizzazione che di estrapolazione, dimostrando che il Deep Learning fornisce risultati più solidi rispetto all'apprendimento automatico standard se applicato ai dati ottenuti in condizioni diverse. Viene inoltre sottolineata l'importanza della complementarità dei dati dei sensori, dimostrando che l'uso di misurazioni basate sull'aspetto può essere utilizzato per prevedere le interazioni ruota-terreno, migliorando la capacità dei rover di esplorazione planetaria di evitare pericoli e rispondere alle condizioni mutevoli. La combinazione di segnali propriocettivi ed esterocettivi viene esplorata per il monitoraggio delle colture nella robotica agricola. La progettazione e lo sviluppo di un sistema multisensore per la ricostruzione 3D di ambienti agricoli è qui dettagliata insieme ai necessari algoritmi di sensor fusion. Inoltre, viene proposto un nuovo algoritmo di apprendimento non supervisionato, Kalman Supervised Network, per la modellazione di sistemi meccanici. KSN utilizza i principi dei filtri Kalman per apprendere continuamente dalle misurazioni dei sensori e produrre un modello del sistema più accurato rispetto a quello integrato nel KF.
Data-driven modelling and estimation of mechanical systems / Vulpi, Fabio. - ELETTRONICO. - (2023). [10.60576/poliba/iris/vulpi-fabio_phd2023]
Data-driven modelling and estimation of mechanical systems
Vulpi, Fabio
2023-01-01
Abstract
This thesis examines the use of machine learning and deep learning in mechanical engineering problems. An underlying theme of the research is the comparison of Convolu- tional Neural Networks and Recurrent Neural Networks with standard machine learning techniques, such as Support Vector Machines, showing the advantages of Deep Learning in modeling complex phenomena where traditional approaches fall short. As an example, it is shown that with the proposed Multichannel Spectrograms, an autonomous robot can classify the traversed terrain with an accuracy of over 90% when using a CNN based on proprioceptive signals. This work addresses both generalization and extrapolation issues, demonstrating that Deep Learning delivers more robust results than standard machine learning when applied to data obtained in varied conditions. The importance of sensor data complementarity is also emphasized showing that the use of appearance- based measurements can be used to predict wheel-terrain interactions, improving the ability of planetary exploration rovers to avoid hazards and respond to changing conditions. The combination of proprioceptive and exteroceptive signals is explored for crop monitoring in agricultural robotics. The design and development of a multi-sensor system for 3D reconstruction of agricultural environments is here detailed together with the necessary sensor fusion algorithms. Additionally, a new unsupervised learning algorithm, Kalman Supervised Network, is proposed for modeling mechanical systems. KSN uses the principles of Kalman Filters to continuously learn from sensor measurements and produce a more accurate model of the system than the one embedded in the KF.File | Dimensione | Formato | |
---|---|---|---|
35 ciclo-VULPI Fabio.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Licenza:
Tutti i diritti riservati
Dimensione
71.01 MB
Formato
Adobe PDF
|
71.01 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.