2026-09-18T21:15:37Zhttps://dspace.mit.edu/server/oai/request

oai:dspace.mit.edu:1721.1/1503052023-04-01T03:47:00Zcom_1721.1_7582com_1721.1_7581col_1721.1_131023

Han, Song Zheng, Jessica Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science 2023-03-31T14:46:30Z 2023-03-31T14:46:30Z 2023-02 2023-02-27T18:43:19.373Z https://hdl.handle.net/1721.1/150305 In recent years, Machine Learning (ML) on edge computing devices has gained attention from both the academic world and industries due to the enormous potential of various applications. Although the advancement in hardware and algorithm optimization techniques have helped accelerate the pace of bringing ML onto edge devices, the resources constraints on such devices remain challenging for ML applications. This work analyzes the feasibility and efficacy of different ML algorithms for two such applications using timeseries data: (1) TinyML for Anomalous Motor Operation Detection, and (2) Estimation of Mean Arterial Blood Pressure (MAP) from ultrasound measurements. In the first application, we explore different algorithms for detecting anomalous fan motor operation on a small microcontroller unit (MCU). Results show that a CNN model can maintain 99% accuracy for anomaly detection even with a small memory footprint of 2.9K parameters (under 6kB of memory). In the second application, we compare different algorithms to optimize the accuracy of MAP estimation from ultrasound data. We find that 1D-CNN and Transformer algorithms using the blood pressure shape and blood flow velocity waveforms can both achieve 8.8mmHg average standard deviation of the prediction error without anthropometric data, and the CNN model can achieve 7.9mmHg when anthropometric data is added as inputs, improving upon a baseline of 9.5mmHg using only anthropometric data. The code for these two projects will be made available at the following links: (1) https://github.com/mit-han-lab/anomaly-detection (2) https://github.com/mit-han-lab/ml-blood-pressure M.Eng. Massachusetts Institute of Technology In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ Machine Learning Applications for Time Series Data: Motor Anomaly Detection and Mean Arterial Blood Pressure Estimation Thesis application/pdf Master Master of Engineering in Electrical Engineering and Computer Science Publication unknown unknown <Publication xmlns="https://www.openaire.eu/cerif-profile/1.1/" id="4427e0a5-c60e-47ae-bf90-e4530d4ed838"> <Type xmlns="https://www.openaire.eu/cerif-profile/vocab/COAR_Publication_Types">http://purl.org/coar/resource_type/c_1843</Type> <Title>Machine Learning Applications for Time Series Data: Motor Anomaly Detection and Mean Arterial Blood Pressure Estimation</Title> <PublishedIn> <Publication> </Publication> </PublishedIn> <PublicationDate>2023-02</PublicationDate> <Authors> <Author> <DisplayName>Zheng, Jessica</DisplayName> <Affiliation> <OrgUnit> </OrgUnit> </Affiliation> </Author> </Authors> <Editors> </Editors> <Publishers> <Publisher> <DisplayName>Massachusetts Institute of Technology</DisplayName> <OrgUnit /> </Publisher> </Publishers> <License>http://rightsstatements.org/page/InC-EDU/1.0/</License> <Abstract>In recent years, Machine Learning (ML) on edge computing devices has gained attention from both the academic world and industries due to the enormous potential of various applications. Although the advancement in hardware and algorithm optimization techniques have helped accelerate the pace of bringing ML onto edge devices, the resources constraints on such devices remain challenging for ML applications. This work analyzes the feasibility and efficacy of different ML algorithms for two such applications using timeseries data: (1) TinyML for Anomalous Motor Operation Detection, and (2) Estimation of Mean Arterial Blood Pressure (MAP) from ultrasound measurements. In the first application, we explore different algorithms for detecting anomalous fan motor operation on a small microcontroller unit (MCU). Results show that a CNN model can maintain 99% accuracy for anomaly detection even with a small memory footprint of 2.9K parameters (under 6kB of memory). In the second application, we compare different algorithms to optimize the accuracy of MAP estimation from ultrasound data. We find that 1D-CNN and Transformer algorithms using the blood pressure shape and blood flow velocity waveforms can both achieve 8.8mmHg average standard deviation of the prediction error without anthropometric data, and the CNN model can achieve 7.9mmHg when anthropometric data is added as inputs, improving upon a baseline of 9.5mmHg using only anthropometric data. The code for these two projects will be made available at the following links: (1) https://github.com/mit-han-lab/anomaly-detection (2) https://github.com/mit-han-lab/ml-blood-pressure</Abstract> <Access xmlns="http://purl.org/coar/access_right" > </Access> </Publication>