Edge-to-Farm Dataset and Code for Wearable Piezoelectric Cattle Behavior Monitoring Using Machine Learning and STFT-Based Deep Learning

This repository contains the dataset, source code, trained models, and embedded implementation files associated with the study: "On-Farm Deployment of a Wearable Piezoelectric Vibration Sensor System for Real-Time Feeding and Rumination Monitoring in Dairy Cows Using Machine Learning and Time–Frequency Analysis" The goal of this work is to develop a low-cost, energy-efficient, and deployable system for continuous monitoring of cattle oro-motor behaviors, specifically Feeding (F), Rumination (R), and Unknown (U), using wearable piezoelectric vibration sensing technology mounted on ear-tag devices. Dataset Description:The dataset consists of real-world vibration signals collected from dairy cows under farm conditions using an ear-mounted piezoelectric sensor connected to an ESP32-WROVER microcontroller. The dataset includes multiple recording sessions and covers three behavioral classes: Feeding, Rumination, and Unknown activities. The raw signals are segmented into fixed-length windows and further processed into statistical, dynamic, and spectral feature representations. In addition, time–frequency representations are generated using Short-Time Fourier Transform (STFT) to capture non-stationary behavioral patterns in the vibration signals. Feature and Data Processing:Extracted features include amplitude-based, statistical, and spectral descriptors such as p90, interquartile range (IQR), skewness, kurtosis, zero-crossing rate, dominant frequency, spectral centroid, spectral entropy, and spectral flatness. These features are used for classical machine learning models (MLP and Random Forest), while STFT representations are used for deep learning-based classification. Machine Learning and Deep Learning Models:Two complementary modeling approaches are provided:1. A lightweight Multilayer Perceptron (MLP) model designed for real-time inference on embedded and edge devices.2. A hybrid deep learning architecture based on STFT + Convolutional Neural Networks (CNN) + Bidirectional Gated Recurrent Units (BiGRU) for high-resolution offline analysis. The MLP model is optimized for deployment on resource-constrained hardware such as ESP32-based systems, while the deep learning model is intended for offline or centralized processing. Embedded Implementation:The repository includes ESP32-compatible implementations and model conversion files (TensorFlow Lite and C header formats). These files are intended for feasibility evaluation of on-device inference under constrained memory and computational resources. Important Note on Deployment:Due to hardware limitations of microcontroller platforms, the full STFT + CNN + BiGRU pipeline is not directly executable on ESP32 devices. Therefore, embedded deployment is achieved using lightweight models only, while deep learning models are used for offline analysis. Applications:This dataset and framework are intended for research in:- Precision livestock farming- Wearable sensor systems- Animal behavior recognition- Edge AI and embedded machine learning- Time–frequency signal analysis Data Structure:The repository includes:- Raw vibration recordings- Processed statistical feature datasets- STFT spectrogram representations- Trained machine learning and deep learning models- Embedded system implementation files- Metadata and documentation This work provides a complete pipeline from low-level sensor acquisition to embedded deployment and deep learning-based behavioral analysis, enabling scalable and real-time monitoring of dairy cattle in real farm environments.