Developing Ultra-lite ML Models for Crash Detection in Noisy Environments

This paper focuses on developing a novel machine-learning (ML) solution on edge-devices, operating in high-noise environments, to detect major faults and events (such as crashes). Every year, thousands of people are subject to bicycle crashes that result from motor-vehicle accidents, poor conditions, etc. A more reliable and on-site anomaly detection will allow for quicker responses in emergencies, helping reduce long-term injuries. Currently, statistical analysis models are being used, but they struggle to detect such anomalies in noisy environments, leading to false alarms. As a solution, an IoT edge device was developed that deploys an ultra-lite ML model to better detect crashes, then notify an emergency contact. To tackle the lack of open-source bicycle crash vibration data, this paper also proposes a novel bicycle crash simulation methodology, utilizing bicycle self-stabilizing properties, to get the data for training. The simple and cost-efficient crash-simulation methodology devised is shown to be rather effective, circumventing the project's cost and other safety-related limitations. The prototype consists of a microcontroller, an IMU, and a GPS. A separate device was also developed to do the data collection and data storage. An ultra-lite ML (<10 KB) was trained, tested, and developed in an iterative process. In a comparison with two statistical analysis models, the ML model had a 130% improvement in accuracy.