Processed Multimodal Dataset for Pig Behavior Recognition Using Wearable Audio and Accelerometer Sensors

This dataset contains processed multimodal data for pig behavior recognition using wearable sensors. The data were collected using a custom electronic collar equipped with a triaxial accelerometer and a digital microphone, under real pig production conditions. The released dataset includes processed 5-second accelerometer windows, corresponding 5-second WAV audio files when available, log-Mel spectrograms generated from the audio windows, behavioral labels, metadata linking accelerometer windows, audio files and log-Mel files, a class dictionary, and a file manifest. The dataset contains 7,892 accelerometer windows distributed into four behavioral classes: lying, eating, walking, and drinking. Each accelerometer window contains 100 samples. The synchronized audio–accelerometer subset contains 6,285 paired samples, each composed of accelerometer data, a corresponding WAV audio file, and a log-Mel spectrogram. The class lying combines the original annotations deitadoac and dormindo. The original annotations correndo, levantando, and deitando were not included in this four-class release. Raw videos are not included due to ethical and institutional restrictions. The dataset was organized to support reproducible machine learning experiments for accelerometer-based, audio-based, and multimodal pig behavior recognition. It can be used to reproduce experiments involving Random Forest models based on handcrafted accelerometer features, convolutional neural networks applied to log-Mel spectrograms, and late-fusion strategies combining audio and accelerometer predictions. The ZIP package includes documentation files describing the dataset structure, variable dictionary, class mapping, metadata files, and validation reports. A SHA256 checksum file is also provided to verify file integrity after download.