Optimizing Behavior Classification Model Performance Through Sensor Data Augmentation Using K-Equidistant Partitioning

In this study, we propose the K-EP (K-Equidistant Partitioning) Augmentation method to enhance the performance of a CNN-based model for classifying pet behaviors using sensor data. The the K-EP Augmentation method involves Partitioning even and odd rows of the time-series data matrix when assuming a binary split, thereby augmenting the data. This approach is based on the research finding that when the sensor data used in previous behavior classification studies satisfies a frequency range of 15 to 40 Hz, it provides sufficient performance for behavior classification models. Additionally, the research suggests that no further significant performance improvement occurs beyond 20 Hz. In our experiments, we aim to validate the performance improvement of the proposed augmentation method compared to various existing augmentation methods using diverse sensor data. The experimental results revealed that the model using only the original data attained an F1-score of 0.801. Models utilizing augmentation techniques demonstrated significant performance improvements K-EP reaching 0.855. Finally, by applying both Overlapping and K-EP techniques simultaneously, an F1-score of 0.903 was achieved.