Deep learning architectures for automated oestrus detection in dairy cattle using multi-sensor collar data: A comparative evaluation of LSTM, CNN, and transformer models under Swedish free-stall housing conditions

Abstract

Can a deep learning model trained on multi-sensor collar data detect oestrus in dairy cattle with sufficient accuracy and lead time to replace or improve upon existing activity-threshold systems? This research compared three deep learning architectures—long short-term memory (LSTM), one-dimensional convolutional neural network (CNN-1D), and Transformer—for automated oestrus detection using data from 284 Swedish Holstein cows fitted with multi-sensor neck collars (triaxial accelerometer, temperature, and rumination microphone) across three commercial free-stall herds in Uppsala County, Sweden, over a 14-month period (September 2022 to November 2023). A total of 1,847 confirmed oestrus events (verified by progesterone profiling) and 18,470 non-oestrus control windows were used for model training (70%), validation (15%), and testing (15%). The Transformer model operating on multi-sensor input achieved the highest performance: 91.7% sensitivity, 96.4% specificity, 92.8% F1-score, and 0.967 AUC-ROC on the held-out test set—significantly outperforming both the CNN-1D (84.2% sensitivity, 86.1% F1, P = 0.003) and LSTM (79.6% sensitivity, 81.4% F1, p<0.001). All multi-sensor models outperformed their activity-only counterparts by 7-12 percentage points in sensitivity. The Transformer maintained sensitivity above 70% when required to detect oestrus at least 8 hours before ovulation, compared with 4-6 hours for the other architectures. These results demonstrate that Transformer-based deep learning applied to multi-sensor collar data provides a practically viable approach to automated oestrus detection that surpasses current commercial activity-monitoring systems.