¶ Spaghetti detection
In X1 seris, we have a USB camera installed in the corner of the printer chamber. With this camera, we can do a lot of wonderful things, e.g. live video stream, timelapse, and most importantly, realtime monotoring of the printing process. Spaghetti detection, among all the others, is the most important function of printing monitoring.
By the word spaghetti, we mean a board range of printing failures, most of which are just like the famous Italian food spaghetti. However, there are also cases where the filament fragments are not so typically spaghetti-ish, but scattered in small pieces. In our spaghetti detection, all these failures are considered the same kind.
¶ How it works
Thanks to the powerful on-chip NPU, we can achive onboard spaghetti detection just inside the printer. Here is the brief procedure:
- Take a frame of the chamber camera every T sec. This period T is configurable. Too small period is unnecessary, since the image won't have enough change in a short time.
- Feed the image to the AI detector running on the NPU. The detector is a state-of-the-art object detection neural network, which finds the bounding box of the potential spaghetti with a confidence value. If the confidence is high enough (the confidence threshold is determined from the sensitivity you choose), it is considered a positive candiate.
- Postprocess the detection results. We accumulate the results for the latest 10 frames. If there are enough positive candiates, and they are spatially consistent, then a spaghetti failure is detected.
To have the detector work properly, there are a few requirements.
- The detector needs to be enabled in the Print Options. You can adjust the sensitivity from {low, medium, high}. If you are off the printer for a relatively long time (e.g. a whole night), you can set the sensitivity to low, so it's less often to pause the printing for small defects.
- The chamber light needs to be open. The spaghettic detector needs a good lighting condition to capture the details of prints. By default we open the light when the first layer starts to print. If the light is shutdown manually, it will not be open back.
¶ How to help
Same as all the deep learning functions, we highly rely on data. If you have been using this function from the very beginning, you'll notice that we have gained a hugh jump in detection accuracy, but still not high enough. There are always false alarms reported from time to time. We really need your help to improve this function, and it's really simple: join the experience improvement program. The image data will be automatically uploaded to cloud if you join the program.