In modern engineering applications, the increasing complexity of systems and the availability of high-frequency sensor data have motivated the development of advanced monitoring, condition recognition and diagnostic frameworks capable of ensuring safety, usage-health monitoring, reliability and adaptability. Physics-based models ensure interpretability and robustness but may fail under non-nominal conditions, whereas data-driven methods, like Machine Learning (ML), excel in pattern recognition but lack physical consistency. This thesis develops two domain-specific frameworks: one for wind turbine monitoring, providing ML-based detection and classification of pitch misalignment anomalies, and another for automotive applications, enabling early classification of road surfaces and tyre types from braking dynamics. Building upon these foundations, the core methodological contribution of this thesis is the development of a domaininformed Mixture of Experts (MoE) framework, which integrates multiple specialized experts within a unified decision-making architecture. Each expert is designed to handle either classification or regression, and operates according to the local properties of the system. The MoE approach allows simultaneous classification and reconstruction of unmeasured signals, while maintaining interpretability and robustness. The proposed methodologies are validated on both real and simulated datasets from the two application domains. Results demonstrate accurate early detection of pitch misalignment in wind turbines, as well as reliable classification of road surfaces and tyre types in automotive scenarios. Overall, the findings confirm the potential of hybrid, domain-informed architectures to improve monitoring, anomaly detection, and predictive decision-making in real-world engineering systems, highlighting their versatility and applicability across structurally different domains.
Nelle moderne applicazioni ingegneristiche, la crescente complessità dei sistemi e la disponibilità di dati da sensori ad alta frequenza hanno motivato lo sviluppo di algoritmi avanzati per il monitoraggio, il riconoscimento delle condizioni operative e la diagnostica, capaci di garantire sicurezza, affidabilità e adattabilità. I modelli basati sulla fisica assicurano interpretabilità e robustezza, ma possono risultare meno accurati in condizioni non nominali, mentre i metodi data-driven, come il ML, eccellono nel riconoscimento di pattern ma generalmente mancano di coerenza fisica. Questa tesi sviluppa due framework specifici per dominio: il primo è dedicato al monitoraggio delle turbine eoliche, fornendo una soluzione basata su machine learning per il rilevamento e la classificazione di disallineamento del passo; il secondo riguarda applicazioni automotive, permettendo la classificazione di superfici stradali e tipi di pneumatici. Sulla base di questi domini, il contributo principale è lo sviluppo di un MoE, che integra più esperti specializzati in un'architettura decisionale unificata. Ogni esperto gestisce classificazione o regressione e opera secondo le proprietà locali del sistema. L'approccio MoE consente classificazione simultanea e ricostruzione di segnali non misurati, mantenendo interpretabilità, robustezza e scalabilità. Le metodologie sono validate su dati reali e simulati dei due domini. I risultati mostrano un rilevamento accurato del disallineamento del passo nelle turbine eoliche e una classificazione affidabile di superfici stradali e pneumatici nelle applicazioni automotive, confermando il potenziale delle architetture ibride nel migliorare monitoraggio, rilevamento di anomalie e decisioni predittive, evidenziandone versatilità e applicabilità a domini differenti.
Learning-based solutions for active monitoring and anomaly detection for advanced industrial applications
Milani, Sabrina
2025/2026
Abstract
In modern engineering applications, the increasing complexity of systems and the availability of high-frequency sensor data have motivated the development of advanced monitoring, condition recognition and diagnostic frameworks capable of ensuring safety, usage-health monitoring, reliability and adaptability. Physics-based models ensure interpretability and robustness but may fail under non-nominal conditions, whereas data-driven methods, like Machine Learning (ML), excel in pattern recognition but lack physical consistency. This thesis develops two domain-specific frameworks: one for wind turbine monitoring, providing ML-based detection and classification of pitch misalignment anomalies, and another for automotive applications, enabling early classification of road surfaces and tyre types from braking dynamics. Building upon these foundations, the core methodological contribution of this thesis is the development of a domaininformed Mixture of Experts (MoE) framework, which integrates multiple specialized experts within a unified decision-making architecture. Each expert is designed to handle either classification or regression, and operates according to the local properties of the system. The MoE approach allows simultaneous classification and reconstruction of unmeasured signals, while maintaining interpretability and robustness. The proposed methodologies are validated on both real and simulated datasets from the two application domains. Results demonstrate accurate early detection of pitch misalignment in wind turbines, as well as reliable classification of road surfaces and tyre types in automotive scenarios. Overall, the findings confirm the potential of hybrid, domain-informed architectures to improve monitoring, anomaly detection, and predictive decision-making in real-world engineering systems, highlighting their versatility and applicability across structurally different domains.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhD_Thesis_SM.pdf
non accessibile
Dimensione
61.5 MB
Formato
Adobe PDF
|
61.5 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/249098