search

TheSpaghettiDetective / obico-server

Obico is a community-built, open-source smart 3D printing platform used by makers, enthusiasts, and tinkerers around the world.
https://obico.io
GNU Affero General Public License v3.0
1.5k stars 299 forks source link

[Feature] Using hot end temperature history to detect anomalies #695

Open rrauenza opened 2 years ago

rrauenza commented 2 years ago

I was reading the prusa Facebook group and saw another blob of death that enveloped the hot end.

It made me wonder… is there an underlying “signal” in the hot end temperature that can be added to the ai training set with the visual training? For example, does the temperature start getting really stable in the hot end when it has a blob of death starting? Or when leaking does it get less stable?

kennethjiang commented 2 years ago

Thank you for this suggestion. I'll love to figure out a way to detect the blob of death.

I guess the tricky part is to collect the data for this problem. My bet is currently we can't find much temperature data during the time when blob of death happens. Short of the ability to retrieve data, the only way is to create data, which by definition will involve manually creating blob of death after blob of death, which sounds like an expensive and nerve-racking endeavor.

Any idea how we can get enough temperature data for the blob of death to start the training process?

rrauenza commented 2 years ago

I was thinking start collecting the temperature data with your user submitted images of failures. Then you've got temperature data classified with success and failures...

I'm wondering if the temperature wiggle history FFT changes with a blob... lower frequency? I think the magnitude wouldn't change due to the PID. But corrective actions by the PID may be less frequent with the extra blob insulation?

At the very least, just start collecting it... so a year from now you have it to poke at.

Maybe simulate it by insulating the hot end and seeing how the PID reacts? See if there is signal in that noise worth pursuing? Obviously using real data later if you think it is detectable...

The hard part may be differentiating a blob vs just the enclosure warming up.

kennethjiang commented 2 years ago

I like your idea of simulating the problem by insulating the hot end and observe if the data come out more or less in line as what the theory predicts. Are you interested in taking a stab at it?

rrauenza commented 2 years ago

No, I don't have the time these days. My thought was to suggest collecting the data from deployments along with images so that it could be analyzed someday.

rrauenza commented 2 years ago

Prusa is putting blob detection in their firmware… with some data analysis in the blog on temperature patterns. You might be interested in this:

https://blog.prusa3d.com/mk3s-3-12-beta-firmware-new-thermal-model-protection-and-blob-detection_71230/

kennethjiang commented 2 years ago

Thank you for the link! Very interesting!