De Donato, Lorenzo (2023) Deep Learning for railway safety and maintenance: methodologies and applications. [Tesi di dottorato]

Anteprima

Testo De_Donato_Lorenzo_36.pdf Download (106MB) | Anteprima

Abstract

This thesis explores the adoption of Deep Learning (DL) for railway safety and maintenance applications, concentrating on solutions leveraging video and audio analysis. After a systematisation of relevant Artificial Intelligence (AI) and railway concepts, the thesis evaluates the opportunities offered by DL in autonomous trains' environment perception and non-intrusive inspections of safety-critical railway assets for maintenance purposes. Then, the thesis focuses on the main challenges concerning the integration of AI into safety-related railway applications, evaluating certifiability and technical issues. The thesis first analyses both aspects methodologically. The concept of "Grade of Intelligence" is introduced for the delineation of high-level guidelines that could support the step-wise integration of AI and DL in the development of intelligent train control systems. Then, a seven-step methodology is presented for the development of DL solutions. Hence, the thesis proceeds with the discussion of two experimental case studies exemplifying the adoption of the aforementioned methodology. The first tackles the problem of anomaly detection, i.e., the identification of unknown objects, on rail tracks. The second deals with the continuous monitoring of level crossings for predictive maintenance. These were chosen for their relevance to safety within the context of study, enabling an exploration of techniques and tools to address – or exploit – specific constraints within the application domain and the lack of suitable datasets for DL models development.

Downloads

Actions (login required)

Modifica documento