TinyML-powered tack-weld detection for robotic welding : A case study towards autonomous industrial manufacturing
Waseem, Hizza (2024)
2024
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2024-12-19
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024121811399
https://urn.fi/URN:NBN:fi:tuni-2024121811399
Tiivistelmä
The combination of robotics and machine learning is transforming digital manufacturing, particularly in robotic welding, where accurate defect detection and monitoring are essential. Tack-welds, which are fundamental in pre-welding assembly, play a critical role in ensuring weld quality. If improperly formed and not identified early, welding over tack-welds can compromise the quality of final weld. Rapid and reliable detection of such welds is therefore crucial to maintain structural integrity.
This research focuses on implementing a TinyML vision-based tack-weld detection system on a Renesas EK-RA6M5 microcontroller integrated with micro-ROS. A lightweight object detection model provided by the Edge Impulse platform was utilized to classify tack-welds in real-time from welding camera images. During initial testing, the model achieved an estimated F1-score of 98.7 %, with a fast inference time of 10 ms and minimal resource consumption (75.2 KB RAM and 78.3 KB Flash). In actual deployment, the system demonstrated an inference time of 80 ms, utilizing 330 KB of RAM and 230 KB of Flash memory.
Despite the increased resource requirements during real-world testing, the system maintained reliable performance, delivering high accuracy and responsiveness. This work demonstrates the feasibility of deploying on-device ML for real-time defect detection, eliminating reliance on cloud-based solutions and reducing latency. The proposed system enhances autonomy in robotic welding processes and serves as a foundation for further advancements in intelligent, edge-based manufacturing systems.
This research focuses on implementing a TinyML vision-based tack-weld detection system on a Renesas EK-RA6M5 microcontroller integrated with micro-ROS. A lightweight object detection model provided by the Edge Impulse platform was utilized to classify tack-welds in real-time from welding camera images. During initial testing, the model achieved an estimated F1-score of 98.7 %, with a fast inference time of 10 ms and minimal resource consumption (75.2 KB RAM and 78.3 KB Flash). In actual deployment, the system demonstrated an inference time of 80 ms, utilizing 330 KB of RAM and 230 KB of Flash memory.
Despite the increased resource requirements during real-world testing, the system maintained reliable performance, delivering high accuracy and responsiveness. This work demonstrates the feasibility of deploying on-device ML for real-time defect detection, eliminating reliance on cloud-based solutions and reducing latency. The proposed system enhances autonomy in robotic welding processes and serves as a foundation for further advancements in intelligent, edge-based manufacturing systems.