Core Concepts
The author introduces CycloWatt, an affordable IoT device revolutionizing cycling power metrics by prioritizing affordability and user-friendliness through cutting-edge technology.
Abstract
CycloWatt is a groundbreaking IoT device that integrates force signals with inertial sensor data using edge machine learning techniques to estimate power accurately. The device operates in low-power mode, offering exceptional battery life and real-time feedback during cycling activities. The paper highlights the importance of cost-effective and accurate cycling power meters, addressing the limitations of existing solutions.
The content discusses the challenges faced by cyclists in adopting power meters due to high costs and deployment complexity. It introduces CycloWatt as a solution that leverages IoT concepts and devices to provide versatile, portable, and energy-efficient power estimation during cycling sessions. The hardware components, data collection setup, data processing stages, and the proposed TinyML model are detailed to showcase the development process.
Furthermore, the experimental evaluation of CycloWatt demonstrates its performance under various conditions such as indoor tests on a roller trainer, outdoor tests, and generalization tests with different riders. The results indicate strong competitiveness compared to previous works in machine learning-based power estimation in cycling. The paper concludes by emphasizing future improvements in mechanical setup, data collection from various riders for enhanced model performance, and the significance of leveraging machine learning for accurate power estimation in cycling.
Stats
Our prototype can estimate power with a Mean Absolute Error (MAE) of only 12.29 Watts (4.1%).
The prototype offers an exceptional battery life of up to 25.8 hours in always-on active mode.
The end-to-end signal inference latency is only 4.33 milliseconds.
The hardware consumes only 0.103 Watts during active mode.
The model requires 270.5 kBytes of memory with full precision accuracy.
The Mean Absolute Error (MAE) for outdoor testing was 15.321 Watts.
Quotes
"Our design emphasizes energy efficiency, operating in a low-power mode that consumes a mere 50 milliwatts."
"With an ultra-low latency of 4.33 milliseconds for data processing and inference."