Automatic monitoring of livestock behavior is an indispensable part of modern farming. It can assist framers in understanding the status of livestock in real-time so as to avoid risks in advance. Compared with other automatic monitoring approaches, such as video or wearable devices, using structural vibration for monitoring can ensure reliability and economy on the premise of including rich information. We propose a pipeline to detect three nursing levels of piglets based on the structural vibration using machine learning methods represented by random forest. Hundreds of features from the time domain and frequency domain with clear physical meaning are extracted at first. The model-based, outcome-oriented feature selection algorithm is implemented then to screen out the sensitive features to the nursing level. It is shown that the statistics of some frequency bands and the local variances in the time domain are essential for detecting the piglet nursing level. After validation, the features extracted through the algorithm still perform well under other common machine learning models. The detection accuracy (F1 score) of the final model with five critical features is up to 0.78 when our approach is deployed in the test set.