A Full End-to-End Deep Approach for Detecting and Classifying Jaw Movements from Acoustic Signals in Grazing Cattle

Monitoring the foraging behaviour of ruminants is a key task to improve their productivity and welfare. During the last decades, several monitoring approaches have been proposed based on different types of sensors such as pressure-based, accelerometers and microphones. Among them, microphones have been one of the most promising options because acoustic signals provide comprehensive information about the foraging behaviour. In this work, a fully end-to-end deep architecture is proposed in order to perform both detection and classification tasks of masticatory events in one step, relying only on raw acoustic signals. The main benefit of this novel approach is the substitution of handcrafted preprocessing and feature extraction phases for a pure deep learning approach, which has shown better performance in related fields. Furthermore, different data augmentation techniques have been evaluated to address the data shortness for models development, typical in this field. The results demonstrate that the proposed architecture achieves a F1 score value of 79.82, which represents an increment close to 18% with respect to other state-of-the-art algorithms. Moreover, the proposed data augmentation techniques provide further performance enhancements, emerging as interesting alternatives in this field.