RATTLE: A Framework for Train Characterization and Identification Through Audio Fingerprinting

Abstract

Train characteristics identification serves a vital role in monitoring the traffic on railway infrastructures and correlating it with their structural response. It traditionally relies on camera-based systems, which raise privacy concerns, have high implementation costs, and perform variably depending on visibility conditions. To address these limitations, we present RATTLE, an IoT framework that uses audio fingerprinting to identify trains, count carriages, and estimate transit speed. Our system combines edge devices for data collection, feature extraction, and processing with cloud infrastructure for aggregation and model training. To validate our approach, we collected a comprehensive dataset of video and audio samples, annotated it with ground truth labels, and used it for evaluation in our experiments. We are also making this dataset openly available to the research community to support further advancements in the field. The framework employs a custom Convolutional Neural Network for processing mel-spectrograms, achieving 96% accuracy in train identification, with a mean absolute error of 0.85 for carriage counting and 1.8 m/s for speed estimation. These results were obtained using only 540 training samples per location, with the model requiring less than 1.63 MB of storage, making it suitable for deployment on resource-limited devices.