JeongHyeon Park, Nammee Moon, Journal of Information Processing Systems Vol. 21, No. 4, pp. 371-379, Aug. 2025  

https://doi.org/10.3745/JIPS.02.0225
Keywords: Behavior Classification, CNN, data augmentation, Sensor data
Fulltext: PDF Full Text PubReader

Abstract

In this study, K-equidistant partitioning (K-EP), a novel data augmentation method, is proposed to address the limitations of sensor data analysis and enhance the performance of behavior classification models. K-EP involves dividing rows of sensor data at equidistant intervals and extracting information from each segment, thereby increasing the size of the dataset by a factor of K. This method is based on the sensor minimum warranted frequency hypothesis, which posits that a sampling frequency of 20–40 Hz provides sufficient data for behavioral classification. The effectiveness of K-EP is validated through three experiments, which involve determining the optimal value of K for K-EP, comparing K-EP with other data augmentation methods, and assessing the added value of K-EP when combined with other methods. The results indicate that K-EP effectively overcomes the quantitative limitations of sensor data and enhances model robustness. It achieves higher F1-scores than existing methods, indicating that it is an effective data augmentation method for sensor-based behavior classification models.

Statistics
Cite this article