Fusing heterogeneous sensor data using artificial intelligence hybrid models for dairy cattle health anomaly detection.

Abstract

This observational study evaluated a hybrid artificial intelligence framework for same day anomaly detection in dairy cows using multimodal sensor data collected under commercial farm conditions. The framework integrated thermal video recorded during routine milking, milking-system measurements, and rumen bolus sensor data from 88 cows. Thermal recordings were processed to extract region-level temperature statistics and deep visual-temporal features using residual convolutional networks and Transformer-based temporal modeling. These thermal features were temporally aligned with production and physiological variables, including somatic cell count, milk yield, electrical conductivity, blood indicators, milk flow, internal body temperature, activity, rumination, and bolus event flags. The fused feature sets were evaluated using several machine-learning classifiers under different modality configurations to assess the contribution of thermal, milking, and bolus data. Model performance depended strongly on modality selection and decision-threshold optimization. The best selected configuration achieved a receiver operating characteristic area under the curve of 0.934 and an F1-score of 0.890, indicating strong discrimination between normal and anomalous health states. Results suggest that production and physiological sensor variables provide strong predictive information, whereas thermal imaging adds complementary value when combined with structured farm data. The proposed framework demonstrates the potential of multimodal artificial intelligence systems to support same day health anomaly detection and decision-support tools for dairy cow health monitoring in real farm environments.