The transportation sector is undergoing a profound transformation driven by the urgency of decarbonization and the growing recognition of the role that active mobility plays in public health. Light Electric Vehicles (LEVs), including e-bikes and electric scooters, are emerging as a strategic solution for sustainable urban mobility within the Italian NRRP/PNRR framework and the MOST National Center for Sustainable Mobility (Spoke 5 - Light Vehicles and Active Mobility). However, the current LEV ecosystem presents unresolved structural challenges: rigid on-board IoT architectures, management of high-frequency data streams on resource-constrained embedded devices, complexity of biometric data collection in real-world conditions, absence of autonomous incident-detection systems, and lack of methodologies for sustainable off-grid charging infrastructures. This thesis addresses these gaps through the systematic transposition of mature industrial paradigms (Edge Computing, containerization, TinyML, and low-power IoT protocols) into the domain of Smart Mobility and Technology for Health. The thesis adopts an architectural approach based on software disaggregation through containerization (Docker), integrated with a multi-access communication infrastructure (LoRaWAN and 5G Standalone). For on-board intelligence, a modular edge node prototype called the "Black Box'' was developed, capable of acquiring high-frequency inertial signals (IMU at 103 Hz) and GNSS data, with subsequent on-device processing and transmission. Two complementary on-device data management strategies were validated: an Autoencoder-based compression model and incremental anomaly detection algorithms (MST and MPT), both optimized for execution on embedded devices through TinyML. Field validation included a biometric data acquisition campaign across all 20 stages of the GiroE-2024 national e-bike tour and simulation-based methodologies for photovoltaic-powered off-grid charging stations. The experimental results confirm the validity of the approach. Lightweight virtualization introduces negligible overhead, with cycle times as low as 2 ms validated in the industrial domain and compatible with the latency requirements of Smart Mobility scenarios. The Black Box prototype achieves 8 hours of operational autonomy, with an average power consumption of 2 W and a peak of 7 W (during data acquisition), and a GPS validity rate of 96.17%. The Autoencoder-based TinyML compression algorithms achieve an 80% compression ratio, with an F1 score close to the 90% reference during subsequent classification validation on automotive OBD data; the MST and MPT anomaly detection algorithms reach an F1 score of 99.84% and 33.67% respectively, with correction error analysis confirming that MST corrections are localized and conservative, while MPT corrections are less localized due to full-sample substitution in the presence of anomalous channels. The energy consumption of a fall detection system was also studied, operating at 12 uA idle current and guaranteeing a battery lifetime exceeding 87 days and a median alert latency of 221 ms. The biometric dataset collected during the GiroE-2024, totaling 180.93 hours and 3749.84 km across 20 stages, represents the first large-scale publicly available physiological dataset from non-professional e-bike riders. The results demonstrate that industrial paradigms can be successfully transposed to mobile embedded platforms, confirming software modularity and distributed Edge intelligence as fundamental enabling requirements for safe, scalable, and maintainable light mobility ecosystems.
Il settore dei trasporti sta vivendo una trasformazione profonda, guidata \\ dall'urgenza della decarbonizzazione e dal crescente riconoscimento del ruolo della mobilità attiva sulla salute pubblica. I Veicoli Elettrici Leggeri (LEV), tra cui e-bike e monopattini elettrici, si affermano come soluzione strategica per la mobilità urbana sostenibile, nel quadro del PNRR e del Centro Nazionale MOST (Spoke 5 - Veicoli Leggeri e Mobilità Attiva). Tuttavia, l'ecosistema LEV presenta criticità strutturali ancora aperte: rigidità delle architetture IoT di bordo, gestione di flussi di dati ad alta frequenza su dispositivi embedded a risorse limitate, complessità della raccolta di dati biometrici in contesti reali, assenza di sistemi autonomi di rilevamento degli incidenti e mancanza di metodologie per infrastrutture di ricarica off-grid sostenibili. Questa tesi colma tali lacune attraverso la trasposizione sistematica di paradigmi maturi nell'ambito industriale (Edge Computing, containerizzazione, TinyML e protocolli IoT a bassa potenza) nel dominio della Smart Mobility e della Technology for Health. La tesi adotta un approccio architetturale basato sulla disaggregazione software tramite containerizzazione (Docker), integrato da un'infrastruttura di comunicazione multi-accesso (LoRaWAN e 5G Standalone). Per l'intelligenza di bordo, è stato sviluppato un prototipo di nodo edge modulare denominato ''Black Box'', capace di acquisire segnali inerziali (IMU a 103 Hz) e dati GNSS, con successiva elaborazione e trasmissione on-device. Due strategie complementari di gestione dei dati sono state validate: un modello di compressione basato su Autoencoder e algoritmi incrementali di anomaly detection (MST e MPT), entrambi ottimizzati per l'esecuzione su dispositivi embedded tramite TinyML. La validazione sul campo ha incluso una campagna di acquisizione biometrica durante le 20 tappe del GiroE-2024 e metodologie di simulazione per stazioni di ricarica fotovoltaiche off-grid. I risultati sperimentali confermano la validità dell'approccio. La virtualizzazione leggera introduce overhead trascurabile, con tempi di ciclo fino a 2 ms validati in ambito industriale e compatibili con le latenze richieste in scenari di Smart Mobility. Il prototipo Black Box raggiunge 8 ore di autonomia operativa, con un consumo medio di 2 W e di picco di 7 W (durante l'acquisizione dati), e una validità GPS del 96,17%. Gli algoritmi TinyML di compressione basati su Autoencoder raggiungono un rapporto di compressione dell'80%, con un F1 score vicino al riferimento del 90% durante la successiva validazione tramite un test di classificazione, validato su dati OBD automotive; gli algoritmi di anomaly detection MST e MPT raggiungono rispettivamente un F1 score del 99,84% e del 33,67%, gli errori di correzione confermano che le correzioni del MST sono localizzate e conservative mentre le correzioni MTP risultano meno localizzate per via della sostituzione dell'intero campione in caso di canali anomali. Si è studiato il consumo energetico di un sistema per la fall detection, operante con 12 uA in idle garantendo un'autonomia superiore a 87 giorni e latenza media di allerta di 221 ms. Il dataset biometrico raccolto durante il GiroE-2024, pari a 180,93 ore e 3.749,84 km su 20 tappe, costituisce la prima raccolta pubblica su larga scala di parametri fisiologici da ciclisti non professionisti. I risultati dimostrano che i paradigmi industriali possono essere trasposti con successo su piattaforme mobili embedded, confermando modularità software e intelligenza distribuita all'Edge come requisiti abilitanti per ecosistemi di mobilità leggera sicuri, scalabili e mantenibili.
Industrial-Grade Edge Computing and IoT Paradigms for Smart Mobility: Enabling Health and Safety Monitoring in Light Electric Vehicles
Gaffurini, Massimiliano
2026
Abstract
The transportation sector is undergoing a profound transformation driven by the urgency of decarbonization and the growing recognition of the role that active mobility plays in public health. Light Electric Vehicles (LEVs), including e-bikes and electric scooters, are emerging as a strategic solution for sustainable urban mobility within the Italian NRRP/PNRR framework and the MOST National Center for Sustainable Mobility (Spoke 5 - Light Vehicles and Active Mobility). However, the current LEV ecosystem presents unresolved structural challenges: rigid on-board IoT architectures, management of high-frequency data streams on resource-constrained embedded devices, complexity of biometric data collection in real-world conditions, absence of autonomous incident-detection systems, and lack of methodologies for sustainable off-grid charging infrastructures. This thesis addresses these gaps through the systematic transposition of mature industrial paradigms (Edge Computing, containerization, TinyML, and low-power IoT protocols) into the domain of Smart Mobility and Technology for Health. The thesis adopts an architectural approach based on software disaggregation through containerization (Docker), integrated with a multi-access communication infrastructure (LoRaWAN and 5G Standalone). For on-board intelligence, a modular edge node prototype called the "Black Box'' was developed, capable of acquiring high-frequency inertial signals (IMU at 103 Hz) and GNSS data, with subsequent on-device processing and transmission. Two complementary on-device data management strategies were validated: an Autoencoder-based compression model and incremental anomaly detection algorithms (MST and MPT), both optimized for execution on embedded devices through TinyML. Field validation included a biometric data acquisition campaign across all 20 stages of the GiroE-2024 national e-bike tour and simulation-based methodologies for photovoltaic-powered off-grid charging stations. The experimental results confirm the validity of the approach. Lightweight virtualization introduces negligible overhead, with cycle times as low as 2 ms validated in the industrial domain and compatible with the latency requirements of Smart Mobility scenarios. The Black Box prototype achieves 8 hours of operational autonomy, with an average power consumption of 2 W and a peak of 7 W (during data acquisition), and a GPS validity rate of 96.17%. The Autoencoder-based TinyML compression algorithms achieve an 80% compression ratio, with an F1 score close to the 90% reference during subsequent classification validation on automotive OBD data; the MST and MPT anomaly detection algorithms reach an F1 score of 99.84% and 33.67% respectively, with correction error analysis confirming that MST corrections are localized and conservative, while MPT corrections are less localized due to full-sample substitution in the presence of anomalous channels. The energy consumption of a fall detection system was also studied, operating at 12 uA idle current and guaranteeing a battery lifetime exceeding 87 days and a median alert latency of 221 ms. The biometric dataset collected during the GiroE-2024, totaling 180.93 hours and 3749.84 km across 20 stages, represents the first large-scale publicly available physiological dataset from non-professional e-bike riders. The results demonstrate that industrial paradigms can be successfully transposed to mobile embedded platforms, confirming software modularity and distributed Edge intelligence as fundamental enabling requirements for safe, scalable, and maintainable light mobility ecosystems.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi corretta Gaffurini.pdf
accesso aperto
Licenza:
Tutti i diritti riservati
Dimensione
12.04 MB
Formato
Adobe PDF
|
12.04 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/380466
URN:NBN:IT:UNIBS-380466