UWSpace is currently experiencing technical difficulties resulting from its recent migration to a new version of its software. These technical issues are not affecting the submission and browse features of the site. UWaterloo community members may continue submitting items to UWSpace. We apologize for the inconvenience, and are actively working to resolve these technical issues.
 

Denoising Autoencoder for Multi-level Random Telegraph Signal Analysis

Thumbnail Image

Date

2023-09-21

Authors

Deng, Bowen

Journal Title

Journal ISSN

Volume Title

Publisher

University of Waterloo

Abstract

Random Telegraph Signals (RTSs) are discrete random fluctuations between different current levels that usually appear in short-channel Metal-Oxide-Semiconductor Field Effect Transistors (MOSFETs). The statistical analysis of RTSs can facilitate the reliability assessment of short-channel MOSFETs. The accurate extraction of parameters in terms of amplitude, level transition, and durations from RTSs is of great importance to statistical analysis. However, background noise like Gaussian white noise or pink noise usually coexists with RTSs, making statistical analysis challenging. Such RTSs with background noise are noisy RTSs, which may contain level fluctuations in both local areas and general trends. To alleviate the impact of background noise on noisy RTSs, RTS reconstruction is usually applied to reconstruct pure RTS traces. The research on RTS reconstruction mainly focuses on models for digitization and parameter extraction, with comparatively limited emphasis on in-depth explorations of RTS denoising. To study the impact of RTS denoising on RTS reconstruction, the research in this thesis explores various RTS denoising models and proposes two Denoising AutoEncoder (DAE) models to enhance the performance of RTS denoising and further improve RTS reconstruction. The research in this thesis generates a synthetic training dataset of multi-level noisy RTSs for DAE training. Four classical denoising models are implemented to compare denoising performance with DAE models on a synthetic evaluation dataset. In addition, three two-level and one multi-level RTS reconstruction models are implemented to validate further that RTS denoising cascading RTS reconstruction can enhance the performance of RTS reconstruction. According to the experimental analysis, DAE models not only have state-of-the-art denoising performance of multi-level RTSs as far as we know. Additionally, these models can dramatically enhance the performance of RTS reconstruction, especially when the pink noise level is high.

Description

Keywords

Random telegraph signal, Deep learning, Background noise, Denoising, Reconstruction

LC Keywords

Citation

URI

http://hdl.handle.net/10012/19904

Collections

Theses
Electrical and Computer Engineering