This repository contains the source code and processed datasets for a deep learning framework designed to multivariate event-based time series classification applied to monitor police firearm usage via embedded sensor fusion. The software features a custom-built Recurrent Neural Network (RNN) library implemented in pure Python/NumPy, enabling the training of lightweight models (Vanilla RNN, GRU, MGRU) suitable for deployment on low-resource embedded systems.

The framework utilizes a multi-objective architecture to process simultaneous inputs from piezoelectric (vibration) and photoelectric (light) sensors. Key capabilities include "Zoneout" regularization, custom Backpropagation Through Time (BPTT), and export functionality to C-style headers for microcontroller integration. The included datasets consist of pre-processed, labelled time-series vectors representing real-world firearm manipulations (shots, reloads, and handling events)

To run the code, the .json files must be placed in a folder named example_dataset located in the same directory as the script.

GRANT INFORMATION:

Grant Agency: UdL-Banc Santander Grant Number: Ajuts UdL-Banc Santander per a la contractació de personal Predoctoral en Formació