This master thesis proposes a novel methodology that mitigates the measurement drift in inertial Micro-Electro-Mechanical Systems (MEMS) mainly due to aging and environ- mental stresses, utilizing an adaptive learning approach. The lack of real-time operational data was worked around by simulating data through a well-thought-out modeling process of MEMS accelerometers. Different movements of sensors have been simulated in various conditions so that the generated simulation data could have a wide coverage over different situations. At the core of the proposed new technique is capturing responses from sensors during particular stimulus and of high precision such that a "ground truth" baseline can be generated. This ground truth will serve for periodic re-calibration through comparison against new responses captured by the sensors under the same stimuli to be able to detect and measure drift. The thesis emphasizes the system architecture designed for seamless integration within the Intelligent Sensor Processing Unit (ISPU) of MEMS sensors or Micro-controller Unit (MCU). As calibration models, tiny neural networks were developed, quantized and vali- dated to ensure that they are lightweight yet robust enough for deployment in resource- constrained environments. These models achieve effective drift compensation without deterioration of the system’s performances and accuracy. The proposed approach en- sures the durability and reliability of MEMS, providing a scalable solution that supports continuous learning and self-calibration while keeping the system operational life signifi- cantly. This approach aligns with the principles of edge AI resulting in enhanced real-time processing and efficient resource utilization as well.
Questa tesi di laurea magistrale propone una nuova metodologia che mitiga la deriva delle misurazioni nei sistemi microelettromeccanici (MEMS) inerziali, dovuta principalmente all’invecchiamento e agli stress ambientali, utilizzando un approccio di apprendimento adattivo. La mancanza di dati operativi in tempo reale è stata superata simulando i dati attraverso un processo di modellazione ben studiato degli accelerometri MEMS. Sono stati simulati diversi movimenti dei sensori in varie condizioni affinché i dati di simulazione generati potessero avere una copertura ampia su diverse situazioni. Il cuore della nuova tecnica proposta è catturare risposte dai sensori durante particolari stimoli e con alta precisione, in modo da poter generare una linea di base di "verità assoluta". Questa verità assoluta servirà per una ricalibrazione periodica mediante il confronto con nuove risposte catturate dai sensori sotto gli stessi stimoli, per poter rilevare e misurare la deriva. La tesi enfatizza l’architettura di sistema progettata per un’integrazione senza soluzione di continuità all’interno dell’Unità di Elaborazione Intelligente del sensori (ISPU) dei sen- sori MEMS o dell’Unità Microcontrollore (MCU). Come modelli di calibrazione, sono state sviluppate, quantizzate e validate piccole reti neurali per garantire che fossero leggere ma abbastanza robuste per essere implementate in ambienti con risorse limitate. Questi mod- elli raggiungono una compensazione efficace della deriva senza deteriorare le prestazioni e l’accuratezza del sistema. L’approccio proposto garantisce la durabilità e l’affidabilità dei MEMS, fornendo una soluzione scalabile che supporta l’apprendimento continuo e l’auto-calibrazione, mantenendo significativamente la vita operativa del sistema. Questo approccio è in linea con i principi dell’edge AI, risultando in un’elaborazione in tempo reale migliorata e un utilizzo efficiente delle risorse.
Continuous IMU-MEMS self-calibration process by means of tiny neural networks
Rezaei, Kiarash
2023/2024
Abstract
This master thesis proposes a novel methodology that mitigates the measurement drift in inertial Micro-Electro-Mechanical Systems (MEMS) mainly due to aging and environ- mental stresses, utilizing an adaptive learning approach. The lack of real-time operational data was worked around by simulating data through a well-thought-out modeling process of MEMS accelerometers. Different movements of sensors have been simulated in various conditions so that the generated simulation data could have a wide coverage over different situations. At the core of the proposed new technique is capturing responses from sensors during particular stimulus and of high precision such that a "ground truth" baseline can be generated. This ground truth will serve for periodic re-calibration through comparison against new responses captured by the sensors under the same stimuli to be able to detect and measure drift. The thesis emphasizes the system architecture designed for seamless integration within the Intelligent Sensor Processing Unit (ISPU) of MEMS sensors or Micro-controller Unit (MCU). As calibration models, tiny neural networks were developed, quantized and vali- dated to ensure that they are lightweight yet robust enough for deployment in resource- constrained environments. These models achieve effective drift compensation without deterioration of the system’s performances and accuracy. The proposed approach en- sures the durability and reliability of MEMS, providing a scalable solution that supports continuous learning and self-calibration while keeping the system operational life signifi- cantly. This approach aligns with the principles of edge AI resulting in enhanced real-time processing and efficient resource utilization as well.File | Dimensione | Formato | |
---|---|---|---|
2024_07_Rezaei_Thesis_01.pdf
non accessibile
Descrizione: Thesis Text
Dimensione
8.48 MB
Formato
Adobe PDF
|
8.48 MB | Adobe PDF | Visualizza/Apri |
2024_07_Rezaei_Executive Summary_02.pdf
non accessibile
Descrizione: Executive Summary Text
Dimensione
1.09 MB
Formato
Adobe PDF
|
1.09 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/10589/221852