VAE-LSTM Data Augmentation for Cattle Behavior Classification Using a Wearable Inertial Sensor

The monitoring of cattle behavior is crucial to understanding their welfare and health status, which can lead to the improvement of output products. The use of neural networks to analyze wearable inertial sensor data attached to cows can be a cost-effective and powerful monitoring method. However, deep learning often requires large-scale datasets, despite its many successes in pattern recognition. For machine-learning-based systems with restricted data sources, data augmentation is an efficient and productive regularization technique. However, the augmentation of time-series motion sensor data remains underexplored. This letter proposes a sequential variational autoencoder long short-term memory (VAE-LSTM), integrating VAE and LSTM, for time-series data augmentation. Our model employs VAE for robust local features and LSTM for long-term correlations. We apply this approach to cattle behavior classification using convolutional neural networks, challenging with limited data. We compare our method with traditional augmentation techniques, achieving significant classification performance improvement from 83.07% to 95.01%.