TY - JOUR
T1 - Leveraging Multi-Image Representations for Improved Time-Series Knowledge Distillation on Wearable Sensor Data
AU - Jeong, Jae Chan
AU - Kim, Jun Soo
AU - Buman, Matthew P.
AU - Turaga, Pavan
AU - Jeon, Eun Som
N1 - Publisher Copyright:
© 2014 IEEE.
PY - 2026
Y1 - 2026
N2 - Wearable sensors are key components of the Internet of Things (IoT), enabling continuous and context-aware monitoring of human activities. As the demand for wearable sensors increases, accurate analysis of time-series data has become critical for reliable activity classification. However, raw wearable sensor signals are often affected by noise and have limited ability to capture contextual and discriminative information. To address these challenges, recent studies have investigated image-based representations derived from time-series data, such as persistence images and Gramian angular fields, which provide richer feature representations for learning. Meanwhile, wearable devices operate under stringent computational and memory constraints, making lightweight models essential for practical deployment. Knowledge distillation (KD) has therefore adopted as an effective approach to transfer knowledge from large, complex models to compact ones suitable for resource-constrained environments. Despite the representational advantages of image-based approaches, their effective integration into KD frameworks for wearable sensor data remains underexplored. In this paper, we analyze how image-based representations contribute to activity classification from wearable sensor signals, and investigate how their incorporation into KD frameworks interacts with model compactness and representation richness. We further extend this analysis to the joint use of multiple image-based representations within KD, investigating their interactions and distillation behaviors. We conduct a comprehensive empirical study and extensive experimental analyses across multiple wearable sensor datasets, evaluating different KD strategies in terms of noise robustness, generalizability, intermediate feature representations, and computational efficiency. The results provide practical design guidelines to design compact and high-performance IoT-based wearable sensing systems.
AB - Wearable sensors are key components of the Internet of Things (IoT), enabling continuous and context-aware monitoring of human activities. As the demand for wearable sensors increases, accurate analysis of time-series data has become critical for reliable activity classification. However, raw wearable sensor signals are often affected by noise and have limited ability to capture contextual and discriminative information. To address these challenges, recent studies have investigated image-based representations derived from time-series data, such as persistence images and Gramian angular fields, which provide richer feature representations for learning. Meanwhile, wearable devices operate under stringent computational and memory constraints, making lightweight models essential for practical deployment. Knowledge distillation (KD) has therefore adopted as an effective approach to transfer knowledge from large, complex models to compact ones suitable for resource-constrained environments. Despite the representational advantages of image-based approaches, their effective integration into KD frameworks for wearable sensor data remains underexplored. In this paper, we analyze how image-based representations contribute to activity classification from wearable sensor signals, and investigate how their incorporation into KD frameworks interacts with model compactness and representation richness. We further extend this analysis to the joint use of multiple image-based representations within KD, investigating their interactions and distillation behaviors. We conduct a comprehensive empirical study and extensive experimental analyses across multiple wearable sensor datasets, evaluating different KD strategies in terms of noise robustness, generalizability, intermediate feature representations, and computational efficiency. The results provide practical design guidelines to design compact and high-performance IoT-based wearable sensing systems.
KW - gramian angular field
KW - knowledge distillation
KW - multimodal representation learning
KW - persistence image
KW - topological data analysis
KW - wearable sensor data
UR - https://www.scopus.com/pages/publications/105051174747
U2 - 10.1109/JIOT.2026.3733757
DO - 10.1109/JIOT.2026.3733757
M3 - Article
AN - SCOPUS:105051174747
SN - 2327-4662
JO - IEEE Internet of Things Journal
JF - IEEE Internet of Things Journal
ER -