Improved feed-forward neural network and multidimensional motion data for accurate recognition of sheep behaviour at multiple fattening stages

Abstract

The behavioural characteristics of Hu sheep are closely linked to their health status. This study focuses on Hu sheep at various fattening stages to explore the role of machine learning in the recognition of multidimensional motion signals and links to sheep behaviour. Multidimensional motion sensors were employed to collect movement data, such as acceleration and angular position from the necks of sheep. Independent sample t-tests were conducted to analyse the temporal distribution differences in sheep behaviours, whilst one-way analysis of variance (ANOVA) was utilised to compare motion data across different behaviours. Factor Analysis (FA) was employed as a dimensionality reduction method whilst retaining relevant motion data and features. An artificial neural network model based on the whale optimisation algorithm (WOA) was established using both the raw and feature dataset, and its performance was compared with other models. The t-test results demonstrated significant differences in the temporal distribution of sheep behaviours at different fattening stages (p < 0.05), while ANOVA revealed significant distinctions among the majority of behavioural data (p < 0.05). The motion data and feature retained through FA effectively preserved key information regarding sheep movement, reducing dimensional redundancy while ensuring the classification efficacy of the model. The integration of WOA with the feature dataset addressed the issue of overfitting and achieved an accuracy of 95.52 %, which is an improvement of 19.04 % compared to traditional models. This research provides a high-precision and high-stability behavioural recognition model for Hu sheep, offering a methodological framework for achieving precise health assessments.