A cow individual milk fat percentage prediction method based on low-rank multimodal fusion of time-series, numerical, and textual data.

Milk fat percentage (MFP) precise prediction is paramount for precision dairy farming and early warning of health issues. However, existing prediction methods primarily rely on single lactation time-series data or statistical models, which often fail to effectively fuse multi-source heterogeneous information and frequently overlook the impact of individual variability. To address current limitations, long short-term memory(LSTM), Stacking ensemble learning,bidirectional encoder representations from transformers(BERT) and low-rank multimodal fusion(LMF) were employed to explore a hierarchical multimodal framework for predicting individual milk fat percentage of dairy cows. This framework not only enables unified modeling and efficient integration of multi-dimensional features (e.g., time-series lactation data, dietary nutrition, gut microbiota abundance, and farm management logs) but also significantly improves the model's predictive accuracy. Meanwhile, through comparative analysis of different modalities and fusion strategies, the importance of critical features is revealed, and the feature combination is further optimized to enhance the framework's practicality and deployment potential. Experimental results show the proposed full-modal fusion mode framework achieves significant improvements over partial fusion models. Specifically, compared with the "Time-series lactation + Dietary nutrition+ gut microbiota" model, the full-modal model reduces root mean square error (RMSE) and mean absolute percentage error (MAPE) by approximately 6.93% and 9.37%, respectively, highlighting the value of pasture management textual information in supplementing management intervention features. Furthermore, relative to the "Time-series lactation+ Dietary nutrition+ Textual Logs" model, the incorporation of gut microbiota features significantly enhances both predictive accuracy and biological interpretability of the full-modal model. Moreover, the hierarchical multimodal fat prediction framework proposed in this study effectively identifies abnormal trends in individual MFP, providing advanced early warnings for metabolic disorders such as subacute rumen acidosis (SARA). This study verifies the effectiveness of multimodal fusion in enhancing prediction performance and provides a scalable pathway balancing accuracy and data acquisition costs, with theoretical and practical value for intelligent livestock management.