Using TinyML to Classify the Flight Phases of an Unmanned Aerial Vehicle

During the development and study of an unmanned aerial vehicle (UAV) the relation between collected data and the flight phase can be a precious information. This work presents a TinyML solution for a real-time classification of flight phases, achieving an accuracy exceeding 80% with an average power consumption of only 0.33 J/s. The implementation employs a MPU6050 module to capture motion data through accelerometer and gyroscope readings, which are then processed via spectral analysis. Post-training quantization techniques are also utilized to achieve the usage of only 3.3 kB of RAM, 24.9 kB of ROM and 20 ms of latency, enabling operation within an ESP32 microcontroller. The motivation for this work stems from the requirements of the PegAzuls AeroDesign team, participating in the SAE AeroDesign competition. This approach demonstrates significant potential for application in various UAV projects, particularly where rapid and precise correlation of flight test data with specific flight phases is essential.