Tese e Dissertação

Tese: Towards Scalable Monitoring Systems: The Benefits of Transfer and Unsupervised Learning in Structural and Process Monitoring Applications

Monitoring systems are essential for structural integrity and industrial process characterization. However, their development faces challenges due to limited labeled data and the need for expert knowledge in feature extraction. This thesis proposes scalable and efficient training frameworks based on Transfer Learning (TL) and Unsupervised Learning (UL) --- aiming to reduce dependency on labeled data, simplify implementation, and increase adaptability across different contexts. Three case studies were conducted. The first introduces a UL framework for estimating tensile stress in aluminum plates using ultrasonic guided waves. Features were extracted using machine, deep, and transfer learning techniques, feeding a k-means model. The best approach achieved 96.00% labeling accuracy, with a 20% error reduction compared to the baseline, and is suitable for real-time execution on in edge computing modules. The other studies apply TL with speech-domain pre-trained models to extract features from accelerometer and acoustic signals for wind turbine blade damage detection and two-phase flow regime classification. The proposed framework achieved 99.60% accuracy in the second study (vs 99.20% for the unisensory baseline), inferring in half the time, and 99.69% in the third (vs 80.00%). The results validate the use of speech models for monitoring tasks involving vibrational and acoustic signals. This thesis provides practical insights into model selection, feature transferability, and computational scalability.

Link da defesa:
https://puc-rio.zoom.us/j/98360632156?pwd=BnCDTxvx6KobrYhvt13xixylngvL2R.1

Ver todos