東華大學圖書館 |

Enabling Machine Learning Tasks in Wearable Cyber-Physical Systems Through Uncertainty Quantification and Signal Processing.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Enabling Machine Learning Tasks in Wearable Cyber-Physical Systems Through Uncertainty Quantification and Signal Processing./
Author:	da Silva, Rafael Luiz.
Published:	Ann Arbor : ProQuest Dissertations & Theses, : 2022,
Description:	161 p.
Notes:	Source: Dissertations Abstracts International, Volume: 84-01, Section: B.
Contained By:	Dissertations Abstracts International84-01B.
Subject:	Wavelet transforms. -
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29176837
ISBN:	9798835548958

Enabling Machine Learning Tasks in Wearable Cyber-Physical Systems Through Uncertainty Quantification and Signal Processing.
da Silva, Rafael Luiz.

Enabling Machine Learning Tasks in Wearable Cyber-Physical Systems Through Uncertainty Quantification and Signal Processing. - Ann Arbor : ProQuest Dissertations & Theses, 2022 - 161 p.

Source: Dissertations Abstracts International, Volume: 84-01, Section: B.

Thesis (Ph.D.)--North Carolina State University, 2022.

This item must not be sold to any third party vendors.

The purpose of machine learning in Cyber-Physical-Systems (CPS) is often the support for automated decision making. However, the application of machine learning to CPS often requires comprehensive studies of the physical phenomenon to guide the choice of a suitable model for the given problem. In this work, several CPS scenarios are explored where signal and image processing, as well as uncertainty quantification can be used to not only enrich the exploratory work for wearable CPS, but actually enable a proper application of machine learning models to solve the problem at hand. This data processing and uncertainty quantification is different than feature engineering, because they are not necessarily utilized directly as inputs of a machine learning model. It is shown that the techniques outlined in this paper, are key factors to understand the underlying physical phenomena.Signal processing techniques are shown as aid to understand the phenomenon of body-rocking, which is an undesired motion that some people do without realizing it. Consequently, such analysis enables the choice of the best methods for designing a wearable body-rocking notification system making use of Inertial Measurement Units (IMU) placed on the upper arm. To further enhance body-rocking detection, it is shown how uncertainty quantification can be used as "quality indicators" for deep learning model's prediction, under the Bayesian Neural Networks framework. Additionally, it is shown how such approach reduces the rate of false positives. In this analysis it explored the effect of such Bayesian approach on the model's capacity. It is shown that the Bayesian framework can provide a performance superior to the original model for the body-rocking detection task. It is also shown how transfer learning between IMU signals from distinct body limbs can reduce the model's prediction variability. Such predictions are the calibrated predicted probabilities of the model ensemble, which are used to calculate the uncertainty in a Bayesian framework.In the realm of ECG monitoring, a method is proposed to transform wearable right arm ECG to chest ECG signals while envisioning future application of machine learning models. This framework enables the portability of current successful models for chest ECG treatment, to seamlessly work indirectly with right arm ECG data.For the application of wearable robotics, a computer vision framework is presented to enable visual context awareness for a robotic lower limb prosthetic. This framework makes use of a camera and an IMU. The signals have been collected and evaluated for establishing the best moment to take a picture. Both the computational resource requirements and performance are analyzed to allow real-time execution of machine learning models. It is shown that Bayesian Neural Networks can reduce power and computational resources usage, by reducing the sampling rate of data acquisition, whenever the model is confident about the upcoming steps.Last but not least, the problem of determining water stress from proximal infrared images and potential fusion with bio-impedance measurements from Maize plants is explored. Infrared reflectance is an indirect indicator of chlorophyll activity and consequently water deficit.

ISBN: 9798835548958Subjects--Topical Terms:

3681479
Wavelet transforms.

Enabling Machine Learning Tasks in Wearable Cyber-Physical Systems Through Uncertainty Quantification and Signal Processing.
LDR:04416nmm a2200349 4500 001 2352250
005 20221118093855.5
008 241004s2022 ||||||||||||||||| ||eng d
020 $a 9798835548958
035 $a (MiAaPQ)AAI29176837
035 $a (MiAaPQ)NCState_Univ18402039428
035 $a AAI29176837
040 $a MiAaPQ $c MiAaPQ
100 1 $a da Silva, Rafael Luiz. $3 3691870
245 1 0 $a Enabling Machine Learning Tasks in Wearable Cyber-Physical Systems Through Uncertainty Quantification and Signal Processing.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2022
300 $a 161 p.
500 $a Source: Dissertations Abstracts International, Volume: 84-01, Section: B.
500 $a Advisor: Lobaton, Edgar.
502 $a Thesis (Ph.D.)--North Carolina State University, 2022.
506 $a This item must not be sold to any third party vendors.
520 $a The purpose of machine learning in Cyber-Physical-Systems (CPS) is often the support for automated decision making. However, the application of machine learning to CPS often requires comprehensive studies of the physical phenomenon to guide the choice of a suitable model for the given problem. In this work, several CPS scenarios are explored where signal and image processing, as well as uncertainty quantification can be used to not only enrich the exploratory work for wearable CPS, but actually enable a proper application of machine learning models to solve the problem at hand. This data processing and uncertainty quantification is different than feature engineering, because they are not necessarily utilized directly as inputs of a machine learning model. It is shown that the techniques outlined in this paper, are key factors to understand the underlying physical phenomena.Signal processing techniques are shown as aid to understand the phenomenon of body-rocking, which is an undesired motion that some people do without realizing it. Consequently, such analysis enables the choice of the best methods for designing a wearable body-rocking notification system making use of Inertial Measurement Units (IMU) placed on the upper arm. To further enhance body-rocking detection, it is shown how uncertainty quantification can be used as "quality indicators" for deep learning model's prediction, under the Bayesian Neural Networks framework. Additionally, it is shown how such approach reduces the rate of false positives. In this analysis it explored the effect of such Bayesian approach on the model's capacity. It is shown that the Bayesian framework can provide a performance superior to the original model for the body-rocking detection task. It is also shown how transfer learning between IMU signals from distinct body limbs can reduce the model's prediction variability. Such predictions are the calibrated predicted probabilities of the model ensemble, which are used to calculate the uncertainty in a Bayesian framework.In the realm of ECG monitoring, a method is proposed to transform wearable right arm ECG to chest ECG signals while envisioning future application of machine learning models. This framework enables the portability of current successful models for chest ECG treatment, to seamlessly work indirectly with right arm ECG data.For the application of wearable robotics, a computer vision framework is presented to enable visual context awareness for a robotic lower limb prosthetic. This framework makes use of a camera and an IMU. The signals have been collected and evaluated for establishing the best moment to take a picture. Both the computational resource requirements and performance are analyzed to allow real-time execution of machine learning models. It is shown that Bayesian Neural Networks can reduce power and computational resources usage, by reducing the sampling rate of data acquisition, whenever the model is confident about the upcoming steps.Last but not least, the problem of determining water stress from proximal infrared images and potential fusion with bio-impedance measurements from Maize plants is explored. Infrared reflectance is an indirect indicator of chlorophyll activity and consequently water deficit.
590 $a School code: 0155.
650 4 $a Wavelet transforms. $3 3681479
650 4 $a Algorithms. $3 536374
650 4 $a Feasibility studies. $3 899065
650 4 $a Prostheses. $3 3683608
650 4 $a Signal processing. $3 533904
650 4 $a Sensors. $3 3549539
650 4 $a Neural networks. $3 677449
650 4 $a Artificial intelligence. $3 516317
650 4 $a Computer science. $3 523869
650 4 $a Electrical engineering. $3 649834
650 4 $a Engineering. $3 586835
650 4 $a Mathematics. $3 515831
690 $a 0800
690 $a 0984
690 $a 0544
690 $a 0537
690 $a 0405
710 2 $a North Carolina State University. $3 1018772
773 0 $t Dissertations Abstracts International $g 84-01B.
790 $a 0155
791 $a Ph.D.
792 $a 2022
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29176837