Clock parameter tracking under temperature variations for time synchronization in wireless sensor networks

Duty-cycled wireless sensor networks rely on tight clock synchronisation to fuse data and schedule radio activity. Yet the frequency of low-cost quartz oscillators drifts non-linearly with temperature, ageing, and supply voltage, so a network that forgoes active time-sync quickly accumulates errors...

Descripción completa

Detalles Bibliográficos
Autor: Monfort Grau, Marc
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/446737
Acceso en línea:https://hdl.handle.net/2117/446737
Access Level:acceso abierto
Palabra clave:Wireless sensor networks
Clocks and watches--Repairing
Sincronització de rellotges
Xarxes de sensors sense fils
Compensació de temperatura
Predicció del desfasament de rellotge
TinyML
LoRaWAN
Clock Synchronization
Wireless Sensor Networks
Temperature Compensation
Clock Skew Prediction
Rellotges--Reparació
Descripción
Sumario:Duty-cycled wireless sensor networks rely on tight clock synchronisation to fuse data and schedule radio activity. Yet the frequency of low-cost quartz oscillators drifts non-linearly with temperature, ageing, and supply voltage, so a network that forgoes active time-sync quickly accumulates errors far beyond the typical ±10 ppm budget. Frequent two-way time-stamp exchanges or temperature-compensated crystals can correct this drift, but they inflate radio airtime, energy consumption, and bill-of-materials cost. The challenge is to extend synchronisation intervals without new hardware or loss of accuracy. This thesis proposes a firmware-only clock-parameter predictor that combines physics-guided curve-fitting with a TinyML neural network to estimate skew in real time, thereby reducing the need for explicit sync messages. Five temperature-indexed skew traces from a LoRa SX1276 mote were recorded between 18 °C and 27 °C. A four-parameter stretched-exponential captures the monotonic decay of each trace, while a cubic term models a transient HVAC-induced overshoot. Synthetic extrapolation balances all traces to 1 500 samples each without extra lab time, and hold-out tests validate the augmentation. A lightweight multilayer perceptron, (in = 2) ¿ 64 ¿ 64 ¿ (out = 1), trained on the augmented dataset, it predicts skew with a mean absolute error of 1.0 × 10¿7 (0.10 ppm), including temperatures withheld during training. The resulting drop in radio airtime yields substantial energy savings without compromising synchronisation accuracy or network reliability. The work enables self-calibrating, sub-ppm clock recovery on battery-powered IoT devices without additional silicon, paving the way for longerlived networks, improved duty cycling, and potential side applications such as node identification via temperature-dependent skew fingerprints.