I have recently aquired quite a lot of filament (abs+ from esun) for a reasonable price. Problem is that while this filament batch is mostly within the usual spec (+-0.05 mm), the diameter oscillates min-to-max roughly every 10 cm. Which manifests on the print like this
You can see the prominent "Z-banding". After figuring out that it was indeed caused by the filament diameter, I went out in search of filament width sensors to try out filament width compensation.
What I found was a bunch of articles on the theory, a lot of designs from a few years back... But absolutely no pictures of "before and after", or any concrete data on the accuracy/repeatability of the designs. Oh, and few comments saying that with modern filaments, it's not worth the effort because if you are within +-0.05 mm, you are fine. Go figure.
In the end I settled on this design, mostly because it looked fleshed out, I could get all the components easily, it was cheap and it was already supported in Klipper. I am using the version with bearings contacting the filament. Total cost is below 3 €.
There is also Thomas Sanladerer's design, the InFiDEL. Requires a bit more hardware, but could probably be equally viable.
I mounted the sensor on my Voron 2.4 at the beginning of the reverse bowden tube (at the filament spool). I wanted to mount it closer to the toolhead, but there are some challenges. Namely, the sensors are sensitive to temperature, and bending of the filament in the reverse bowden can change the readings by putting differing pressure on the lever arm at different positions of the toolhead. Best compromise would be to mount the sensor at the very last point before the filament enters the chamber.
Next up, tuning the raw sensor values. I used a 1.5 mm allen key and a 2 mm drill bit (precise size verified with a micrometer), as is recommended in the guide. Fortunately, when filament was inserted, the values were corresponding with my measurements with an error on the order of units of micrometers. Pretty impressive. I tried multiple types of filament, all worked similarly well. If you have problems with this step, maybe consider the pre-tension spring being too strong and squishing the filament - I had to shorten mine a bit with snips to get good results. Also think about how the filament beds coming from the spool - ideally you want to measure it "from the sides" where it does not curve, since curving filament can put additional force on the measuring arm and skew the results.
Then I had to tune the distance between the sensor and nozzle. I got a rough measurement by pushing filament all the way in, snipping it at the sensor, pulling it out and measuring the length. This came out to roughly 128 cm. Just to be sure, I disconnected my hotend, made a contrasting mark at the sensor and ordered the extruder to push filament until the marked piece came out. This came out to 128.5 cm from sensor to nozzle end, or roughly 126.5 cm from sensor to melt zone (Dragon HF).
To be sure, I wanted to try a few distances to see how sensitive the algorithm is to this parameter. Since there is no "tuning tower"-style procedure for fine tuning, I elected to brute-force the problem by printing a 3-wall, 0-infill, 80x80x20 mm cube (with a 1 mm brim for adhesion) for every 10 mm of the sensor-nozzle distance, starting at 128.5 cm and going lower with every iteration. Such a cube consumes roughly 2.5 meters of filament, so any compensation should start happening roughly in the middle of the print.
Shown below are 3 cubes, from the left:
- Control. No compensation.
- Compensation on from a printer reset (clean filament width data buffer). Actual compensation kicks in roughly in the middle of the print.
- Compensation on, no reset (filament width data buffer full from the previous print).
Keep in mind that this is the harshest possible lighting I could provide, direct overhead white LEDs. Same thing, slightly more favorable lighting
These 3 were taken with a distance parameter of 124.5 cm, but really anything from around 123 to 128 cm was usable. Other tests with Prusament ASA gave me the best subjective result at 126.5 cm, which is the setting I ended up using globally. Since it is a distance between the sensor and the start of the melt zone, it makes sense to me.
Here is the print from the first image in this repo, plus the same thing with comp enabled
It's far from perfect, but I am happy with it
Klipper's built-in compensation code allows logging the measured diameter. This is a bit clunky, as it just pastes the measurements into the console at the measured intervals (I use 1 mm, lowest possible distance do not use 1 mm, see appendix B), which then gets saved to klippy.log. From there, you can extract the data and plot the filament diameter...
...yikes. This correlates well with the observed effects (wide band of overextrusion followed by a thin band of underextrusion). Notably, this is still within the "standard" tolerance of +-0.05 mm, so the filament technically passed the basic QC requirement. Varying the diameter this much over such short distances does produce visible artefacts on the print though, as demonstrated. And even a cheap 3€ mechanical filament sensor can be used to fix the issue.
Note - I would not completely trust the X axis. I made a few prints that required 5m of filament but only got ~2200 log entries (I use 1 mm measurement_interval not anymore, standard 10 mm is good enough). I have yet to determine whether the logging is time based, time-limited, or what else is going on, so just... Take the X axis with a grain of salt See appendix A.
I have been using the sensor for a grand total of 3 days at this point. No idea if these results are reliable in day-to-day printing as of yet.
Small update - It's been a few months and I can say that this compensation has given me zero issues. Everything just works, prints look great once the filament compensation kicks in (if the filament diameter is crap, otherwise there is not much change).
I have no idea about what accuracy/precision/repeatability to expect from the sensor, although the resolution seems to be about 1 micron. With good and consistent filaments, it is entirely possible that using the compensation would introduce more errors into the print than would otherwise be there, so use at your own risk*. Still, it makes my stash of esun ABS+ printable, so I am happy with it so far.
Also be aware that you might have to retune your EM with this mod. Since this process messes with flow rate, the slicer-side EM might be a bit different than before, i.e. if your filament is really consistent, but has a nominal mean diameter of 1.72 mm, your prints will look great, but your tuned EM might be higher than usual. The filament width compensation, however, expects 1.75 mm, so it will try to overextrude when it sees 1.72 mm. As a result, your slicer-side EM would actually need to be a bit lower than expected in this hypothetical example to get good prints.
It would also be beneficial to have at least 2 sensors working in unison to get a better measurement of the filament dimensions. Klipper is not ready for it yet though - maybe someone can add that into the existing module?
Also, as the sensors work in analog, they are susceptible to electrical noise - a lowpass filter near the board input pins might be advisable. Or go with a digital solution in the first place.
I'll try to keep an up-to-date chart with various filament width logs here. Click on the chart to get an interactive live version.
I found that the logs are a bit wrong - on the X axis, the units are definitely not mm, they are more like 2.533 mm for each sample. Filaments are still comparable within this chart, just note that the trends you see in them are longer than appears at first glance. (Reason for this is in Appendix A)
So far we have:
- ESUN ABS+ Black, new batch
- Prusament PLA Mystic Green
- ESUN ABS+ Natural, new batch
- ESUN ABS+ Fire Engine Red, old batch
- Prusament ASA Galaxy Black
So what if you only run good filament? Do you need a filament width sensor? This is a test piece made from Prusament Galaxy Black ASA - you can see in the chart above that this filament has good diameter tolerances and stability. What does the compensation do with such a filament? You be the judge:
In my opinion, not much, really. I can see a small improvement from halfway up if I squint really hard, but that could easily be chalked up to crappy EM tuning on my part (which the compensation fixes). Don't expect miracles, after all, this is still a robotic hot glue gun. At least the compensation doesn't seem to degrade the quality, which is a nice peace of mind.
I noticed that for the test boxes, I was getting roughly 2-3x less samples than expected. This is because the algorithm can't update the values (and report them to console, which happens in the same function) while a gcode line is executing. If I extrude less than or equal amount of filament with
M83
G1 E<distance> F<speed>
compared to the defined measurement distance, I get perfectly fine log reports. i,e, set measurement distance to 10, do 10x 1mm extrudes, get 1 log report. Do 2x 5mm extrudes, get 1 log report. Do 1x 10mm extrude, get 1 log report. However, if I overshoot the measurement distance, e.g. with 1x 25mm extrude, I only get 1 log report even when I was expecting 2. Funnily enough, when this 25mm extrude finishes, I need to do another 10x 1mm extrudes to get another log report.
Retracts seem to work fine too, if I do 3x 25mm retract, then extrude 7x 10mm and 1x 5mm, I need to do another 10 mm to get the next log report It also seems that the algorithm correctly remembers the last reported position, since extrude 50 mm (gives 1 log report), retract 50 mm and then extrude 5x 10 mm does not give me new log reports - but the next 10 mm extrude does.
This is the reason why my test boxes get 2-3x less samples than expected - each wall is 80 mm long, each line is 0.2 mm high and 0.38 mm wide. 80*0.2*0.38 = 6.08 mm^3, which is roughly 2.5333 mm.
I noticed that while all of my test boxes came out exactly as expected, if I started a real print, the filament width compensation algorithm broke. I did a multitude of tests and this behavior was consistent - consecutive test boxes came out great, real print, next box was completely trash, as if the algorithm threw out random values. Every time, regardless of how many test boxes were printed before the real print.
I thought maybe my print speeds were overwhelming the algorithm, so I printed the real print capped at 1 mm/s (2.4 mm^3/s). No dice, same behavior.
Then I tried completely disabling retractions and pressure advance along with retractions in custom macros... Nope, still broken.
Finally I set the measurement distance back to the stock value, 10 mm - problem disappeared. Tested with an additional real print and the box after it came out great.
I have no idea what causes this behavior as of yet, or whether the stock 10 mm measurement distance works or just breaks down much slower.