Micro-Electro-Mechanical Systems (MEMS) Inertial Measurement Units (IMUs) have revolutionized sensor technology, enabling widespread adoption in smartphones, wearable devices, autonomous driving, and robotic systems. However, these compact sensors suffer from measurement imperfections, drift over time, and high noise levels, particularly affecting angular velocity measurements from gyroscopes. Traditional calibration methods, typically employing linear mathematical models with few parameters, struggle to capture the complex nonlinear error patterns inherent in MEMS IMUs. This work focuses specifically on the gyroscope calibration aspect, which represents a simplification of the broader MEMS IMU calibration problem. This thesis presents a study on the calibration behavior of a novel data-driven denoising method focused on the gyroscope component of MEMS IMU data. Leveraging Physics-Informed Neural Networks (PINNs), the method integrates physical constraints through differential equations that relate angular velocity measurements to orientation data. Building upon the Physics-Informed Learning for Denoising Technology (PILOT) framework, the proposed approach operates solely with gyroscope data during deployment, unlike existing methods that require multiple sensor inputs, making it suitable for resource-constrained environments. The core methodology involves a two-phase training process: a reconstruction phase that learns to remove artificially injected noise, followed by a physical alignment phase that enforces consistency between integrated gyroscope data and ground truth orientations. A significant contribution of this work is the creation of the ST dataset, comprising controlled robotic turntable movements with precise ground truth measurements, along with a semi-supervised labeling method. The dataset enables comprehensive evaluation of the model's performance under controlled conditions, complementing experiments conducted on the established OxIOD dataset. Experimental results demonstrate that the proposed model exhibits dual capabilities depending on dataset characteristics: it functions as a denoising mechanism for complex, variable datasets (achieving a 19% improvement in average quaternion angle distance on OxIOD), while displaying calibration properties for controlled datasets (achieving a 78% improvement on the ST dataset). This adaptive behavior reveals the model's ability to correct both stochastic noise and systematic biases, effectively bridging the gap between traditionally separate denoising and calibration tasks. The findings indicate that under certain conditions, particularly with controlled motion patterns, the denoising model demonstrates emergent calibration properties, specifically the ability to correct constant errors (bias) in gyroscope data. This dual functionality, combined with the model's lightweight architecture, makes it well-suited for real-world deployment in IoT applications and embedded systems where both denoising and calibration are critical for accurate motion sensing.
Le Micro-Electro-Mechanical Systems (MEMS) Inertial Measurement Units (IMUs) hanno rivoluzionato la tecnologia dei sensori, permettendone un'adozione diffusa in smartphone, dispositivi indossabili, guida autonoma e sistemi robotici. Ciononostante, questi sensori compatti soffrono di imperfezioni di misura, deriva nel tempo e alti livelli di rumore, che influenzano particolarmente le misurazioni di velocità angolare ottenute dai giroscopi. I metodi di calibrazione tradizionali, che tipicamente impiegano modelli matematici lineari con pochi parametri, faticano a catturare i complessi schemi di errore non lineari intrinseci nei MEMS IMUs. Questo lavoro si concentra specificamente sull'aspetto della calibrazione del giroscopio, che rappresenta una semplificazione del più ampio problema di calibrazione dei MEMS IMUs. Questa tesi presenta uno studio sul comportamento di calibrazione di un nuovo metodo data-driven di denoising focalizzato sul componente giroscopico dei dati MEMS IMU. Sfruttando le Physics-Informed Neural Networks (PINNs), il metodo integra vincoli fisici attraverso equazioni differenziali che collegano le misurazioni della velocità angolare ai dati di orientamento. Basandosi sul framework Physics-Informed Learning for Denoising Technology (PILOT), l'approccio proposto opera esclusivamente con i dati del giroscopio durante il deployment, a differenza dei metodi esistenti che richiedono input da più sensori, rendendolo adatto ad ambienti con risorse limitate. La metodologia centrale prevede un processo di training in due fasi: una fase di ricostruzione che impara a rimuovere il rumore artificialmente iniettato, seguita da una fase di allineamento fisico che impone la coerenza tra i dati integrati del giroscopio e le orientazioni di ground truth. Un contributo significativo di questo lavoro è la creazione del dataset ST, che comprende movimenti controllati su un tavolo rotante robotico con misurazioni di ground truth precise, insieme a un metodo di etichettatura semi-supervisionato. Il dataset consente una valutazione completa delle prestazioni del modello in condizioni controllate, integrando gli esperimenti condotti sul consolidato dataset OxIOD. I risultati sperimentali dimostrano che il modello proposto mostra capacità duali a seconda delle caratteristiche del dataset: funziona come meccanismo di denoising per dataset complessi e variabili (raggiungendo un miglioramento del 19% nella distanza media dell'angolo tra quaternioni su OxIOD), mentre mostra proprietà di calibrazione per dataset controllati (raggiungendo un miglioramento del 78% sul dataset ST). Questo comportamento adattivo rivela la capacità del modello di correggere sia il rumore stocastico sia i bias sistematici, colmando efficacemente il divario tra i compiti tradizionalmente separati di denoising e calibrazione. I risultati indicano che, in determinate condizioni, in particolare con pattern di movimento controllati, il modello di denoising manifesta proprietà emergenti di calibrazione, in particolare la capacità di correggere errori costanti (bias) nei dati del giroscopio. Questa doppia funzionalità, unita all'architettura leggera del modello, lo rende particolarmente adatto per il deployment nel mondo reale in applicazioni IoT e sistemi embedded dove sia il denoising sia la calibrazione sono critici per una misurazione accurata del movimento.
Calibration behavior in MEMS IMU gyroscope denoising model
Corradini, Claudio
2024/2025
Abstract
Micro-Electro-Mechanical Systems (MEMS) Inertial Measurement Units (IMUs) have revolutionized sensor technology, enabling widespread adoption in smartphones, wearable devices, autonomous driving, and robotic systems. However, these compact sensors suffer from measurement imperfections, drift over time, and high noise levels, particularly affecting angular velocity measurements from gyroscopes. Traditional calibration methods, typically employing linear mathematical models with few parameters, struggle to capture the complex nonlinear error patterns inherent in MEMS IMUs. This work focuses specifically on the gyroscope calibration aspect, which represents a simplification of the broader MEMS IMU calibration problem. This thesis presents a study on the calibration behavior of a novel data-driven denoising method focused on the gyroscope component of MEMS IMU data. Leveraging Physics-Informed Neural Networks (PINNs), the method integrates physical constraints through differential equations that relate angular velocity measurements to orientation data. Building upon the Physics-Informed Learning for Denoising Technology (PILOT) framework, the proposed approach operates solely with gyroscope data during deployment, unlike existing methods that require multiple sensor inputs, making it suitable for resource-constrained environments. The core methodology involves a two-phase training process: a reconstruction phase that learns to remove artificially injected noise, followed by a physical alignment phase that enforces consistency between integrated gyroscope data and ground truth orientations. A significant contribution of this work is the creation of the ST dataset, comprising controlled robotic turntable movements with precise ground truth measurements, along with a semi-supervised labeling method. The dataset enables comprehensive evaluation of the model's performance under controlled conditions, complementing experiments conducted on the established OxIOD dataset. Experimental results demonstrate that the proposed model exhibits dual capabilities depending on dataset characteristics: it functions as a denoising mechanism for complex, variable datasets (achieving a 19% improvement in average quaternion angle distance on OxIOD), while displaying calibration properties for controlled datasets (achieving a 78% improvement on the ST dataset). This adaptive behavior reveals the model's ability to correct both stochastic noise and systematic biases, effectively bridging the gap between traditionally separate denoising and calibration tasks. The findings indicate that under certain conditions, particularly with controlled motion patterns, the denoising model demonstrates emergent calibration properties, specifically the ability to correct constant errors (bias) in gyroscope data. This dual functionality, combined with the model's lightweight architecture, makes it well-suited for real-world deployment in IoT applications and embedded systems where both denoising and calibration are critical for accurate motion sensing.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis.pdf
non accessibile
Descrizione: Documento di tesi
Dimensione
4.12 MB
Formato
Adobe PDF
|
4.12 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/10589/240149