VibraFarrow: Pig Farrowing Time Prediction Using Ambient Floor Vibrations

Farrowing, the onset of parturition in mother pigs (i.e., sows), is a high-risk period for both the sow and her newborn piglets. Early and accurate prediction of farrowing time, along with monitoring indicators such as vital signs and pre-farrowing behaviors, enables timely assistance and can lead to lower stillbirth rates. However, existing methods have limitations: camera-based systems require constant lighting that disrupts pigs’ circadian rhythms, while wearable sensors can cause discomfort to pigs and are prone to be damaged.We introduce VibraFarrow, a novel non-intrusive vibration sensing system that models pig-induced floor vibration data collected from the built environments to predict farrowing time. Specifically, VibraFarrow predicts whether a sow will farrow within the next 20 hours, enabling contact-free, long-term tracking of farrowing-related behaviors for pigs. The key challenge in developing VibraFarrow is the uncertainty of pre-farrowing behaviors embedded in the long-term vibration time series. First, pre-farrowing behaviors exhibit highly variable timing and duration, occurring intermittently and mixed with ambient noises, which complicates feature extraction over time. Second, the types and patterns of pre-farrowing behavior are uncertain, leading to unreliable predictions using a fixed set of explicit activities. To address the first challenge, VibraFarrow introduces Hierarchical Adaptive Window Selection (HAWS), a hierarchical method that adaptively selects time windows ranging from hours, minutes, to seconds and extracts farrowing-related features from long-term ambient vibration data. Furthermore, to overcome the uncertainty of pre-farrowing behavior patterns and types, VibraFarrow allows flexibility in extracting implicit indicators that are representative of pre-farrowing behaviors for specific time windows through unsupervised clustering. Finally, VibraFarrow fuses the implicit behavior indicators and other expert-defined features (e.g., heart and respiration rates) extracted by HAWS for farrowing time prediction.We deployed our system on a real-world farm for seven months, monitoring 18 farrowing events and collecting 384 hours of vibration data. The system achieved a weighted F1-score of 0.735, surpassing the baseline by up to 20%. This improvement demonstrates floor vibration sensing as an effective, non-intrusive approach for farrowing prediction, with broader implications for occupant health and wellness management in built environments.