Um estudo sobre modelos de redes neurais para identificação de estresse em vocalizações de gado bovino

Abstract

Beef cattle farming plays a fundamental role in the Brazilian economy, being one of the main sources of income for the country. During various stages of production, such as veterinary procedures, weighing, and transportation, animals are often subjected to different levels of handling, each with its potential to cause stress in the animals. Stress developed by the animals can have a significant impact on the properties of the meat, ranging from reducing its quality to making it unsuitable for human consumption. Additionally, poorly conducted handling can result in injuries to the animals, ranging from minor bruises to situations that lead to the death of the animal, resulting in economic losses for producers and slaughterhouses, as well as the suffering of the animals themselves. The aim of this study was to evaluate different approaches to identifying the occurrence of stress in animals through the sounds emitted during handling. The evaluation was carried out in two ways: (i) the implementation of different neural network architectures, including MLP (Multilayer Perceptron), LSMT (Long Short-Term Memory), and CNN (Convolutional Neural Network), using MFCC (Mel Frequency Cepstral Coefficients) for acoustic feature extraction, and (ii) the statistical analysis of acoustic parameters of vocalizations, such as fundamental frequency (F0), spectral formants (F1 - F4), jitter, shimmer, harmonicity and intensity. The results showed that convolutional neural networks achieved better stress classification rates (F1-score = 98.76%), followed by MLP networks (F1-score = 90.07%) and LSTM (F1-score = 85.61%). In the acoustic analysis, significant differences (p < 0.001) were observed between the parameters of stress and non-stress vocalizations in most cases, reinforcing the feasibility of using acoustic characteristics to identify the emotional state of animals during handling.