Improved composite deep learning and multi-scale signal features fusion enable intelligent and precise behaviors recognition of fattening Hu sheep

The integration of artificial intelligence and advanced sensing technologies can improve the intelligence and precision level of livestock management. This study focuses on fattening Hu sheep as the object of study, and aims to assess the effectiveness of integrating multi-scale biological signals with improved composite deep learning model in identifying and classifying behaviors of fattening Hu sheep. The multi-scale biological signals were collected using the respiratory sensor and the multi-dimensional posture sensor (composed of an accelerometers, gyroscope, and magnetometer), and then, after data processing, extracted signal features and used the dimensionality reduction method of principal component analysis (PCA). Attention-based particle swarm optimized convolution and long short-term memory (APSO-CALM) model was developed using the feature fused dataset, and its performance was compared with other models. The results showed that: (1) The multi-scale biological signals were analyzed and categorized into five distinct behaviors based on experimental records: feeding, rumination, mating, free movement and running. Each of these behaviors exhibits unique characteristics in their signal images. (2) PCA was utilized to reduce the dimensionality of the feature fused dataset of the multi-scale biological signals, preserving principal components with a cumulative contribution rate of 98 %. Among all components of the first and second contribution rates, except for a few individuals, there are significant differences (P < 0.05) between the data of different behaviors of the same component. (3) The improved composite deep learning model, APSO-CALM, demonstrates significant advantages over single models in behavior recognition. Its accuracy, precision, recall, and F1 score are 95.0 %, 94.8 %, 94.5 %, and 94.6 %, respectively. By utilizing the APSO-CALM model, the drawbacks of individual models are mitigated, enhancing overall performance and overcoming the limitations of single model applications. This study effectively identified five behaviors of fattening Hu sheep, providing theoretical and practical basis for intelligent and precise management of fattening Hu sheep.