Real-Time FDM Defect Detection | AI/ML for Modern Manufacturing

For this project, I was tasked with training a computer vision model from Ultralytics called YOLO (v9 and v11 were used) to identify and characterize defects in FDM prints.

Training process: Obtaining a dataset was the biggest challenge of this project. Roboflow was a great website for finding large datasets, but the quality of the images and labeling varied greatly. The first models I trained ran into a lot of issues because of this, so I ended up combining images from different datasets using a Python script. I found that using large batches reduced noise in the first few epochs of training, and using larger epochs with a low learning rate, along with reasonable patience, helped with over/under-fitting the data.

Impact & Reflection: This project highlighted the biggest limitation in vision models, which is curating training data. Without clear images, proper bounding boxes for labels, and a reasonable split between learning and validation data, the model was not able to recognize defects with a high enough confidence.

Key Skills & Concepts Used: Google Colab, Python, YoloV11, Vision Model, FDM, Roboflow