Preprocessing Strategies for Animal Behavior Classification Using Inertial Sensors: Effects of Filtering, Normalization, and Data Representation

Automatic classification of livestock behaviors, including feeding, rumination, standing, lying, walking, and drinking, using wearable accelerometers has become an important tool in precision livestock farming (PLF). However, the influence of preprocessing strategies on classification performance and computational efficiency remains poorly understood. This study investigated the effects of filtering, normalization, data representation, and machine learning algorithms using five accelerometer datasets comprising 3,533,974 records collected from different livestock species and sampling frequencies. Four filtering strategies, three normalization methods, two data representations, and four machine learning algorithms were evaluated using a standardized pipeline. Model performance was assessed using weighted F1-scores together with statistical and computational analyses. The absence of filtering achieved the highest average performance, reaching a weighted F1-score of 0.773 with Random Forest, whereas high-pass filtering consistently reduced performance (minimum average F1 = 0.605 across sampling frequencies) while increasing computational cost. Z-score standardization improved the performance of scale-sensitive algorithms, increasing SVM performance by up to 8.6% compared with no normalization. Feature-based representations provided greater stability for conventional machine learning models, whereas raw signals generally benefited the 1D-CNN. These findings demonstrate that preprocessing strategies should be selected according to the learning algorithm, signal characteristics, and computational constraints rather than applied as universal procedures, providing methodological guidance for the development of efficient livestock behavior monitoring systems.