TinyML pipeline for efficient crack classification in UAV-based structural health inspections.

Area of Science:

• Embedded Systems
• Machine Learning
• Structural Health Monitoring

Background:

• UAVs with vision sensors are crucial for infrastructure inspection.
• Cloud-based crack classification faces limitations in bandwidth, power, and privacy.

Purpose of the Study:

• To develop a self-contained Tiny Machine Learning (TinyML) pipeline for onboard crack classification.
• To optimize the pipeline for low-power microcontrollers and evaluate its performance.

Main Methods:

• Utilized MobileNetV1x0.25 as a baseline for a TinyML pipeline on an STM32H7 microcontroller.
• Compared two image preprocessing strategies and four model compression techniques (PTQ, QAT, pruning, weight clustering).
• Evaluated the full measurement pipeline including image capture, preprocessing, and inference.

Main Results:

• Achieved an F1-score of 0.938, an 11.4% improvement over state-of-the-art.
• Required only 2.9 MB RAM and 309 KB flash, with 461.6 ms latency and 623.16 mJ energy cost per inference.
• Demonstrated minimal impact on UAV flight time (4%) compared to power-hungry platforms (24%).

Conclusions:

• The developed TinyML pipeline offers a practical balance of accuracy, resource efficiency, and energy consumption for UAV-based SHM.
• This work establishes a reproducible benchmark for on-device crack classification in resource-constrained environments.
• Advances the feasibility of real-time, on-board structural health monitoring using embedded AI.