We all have our preferences when it comes to soldering irons, and for [Marius Taciuc] the strongest of them all is for a quick heat-up. It has to be at full temperature in the time it takes him to get to work, or it simply won’t cut the mustard. His solution is a temperature controlled iron, but one with no ordinary temperature control. Instead of a normal feedback loop it uses a machine learning algorithm to find the quickest warm-up.
The elements he’s using have a thermocouple in series with the element itself, meaning that to measure the temperature the power must be cut to the element. This duty cycle can not be cut too short or the measurements become noisy, so under a traditional temperature control regimen there is a limit on how quickly it can be heated up. His approach is to turn it on full-time for a period without stopping to measure the temperature, only measuring after it has had a chance to heat up. The algorithm constantly learns how long to switch it on to achieve what temperature, and is able to interpolate to arrive at the desired reading. It’s a clever way to make existing hardware perform new tricks, and we like that.
He’s appeared on these pages quite a few times over the years, but perhaps you’d like to see the first version of the same hardware. Meanwhile watch the quick heat up in action with a fuller explanation in the video below.
Pretty good toy example for machine learning.
My soldering station needs ~10 seconds to heat up, and as far as I can see it does the same thing – full on blast for a while and then goes into the loop. If i remove the tip, i can see the heating element reaching red glow for a few seconds at startup. I am guessing that when you know precisely all the parameters like the heating element and supply you can forget the the machine learning and it should be an almost linear relationship between the time you turn it on full blast and the desired temperature.
It is not a linear relationship as the resistance of the heater and heat loss both have a temperature dependence.
I would bet that the amount of off time is or can be tweaked to a small enough % that t really doesn’t matter in real life. During the learning phase, the machine learning controller is sub-optimal.
It is a matter of obtaining a temperature vs time profile using the original heat-off-read temperature cycle, factor in the tiny amount of gaps. The profile let you calculate the time between starting point to desired end point. That should be pretty close anyway on first try.
The Pace ADS200 manages as good performance as this without machine learning. Not wanting to be a hater, but i can’t stand Machine learning in everything.
Most integrated cartridge-type soldering irons do this. Full blast for a set time and then chill out.
Very nice! If machine learning is the same technology that guesses what I want to type and fills in the (wrong) word for me, then I agree with Alex. But what you have is a system that learns once and then follows those rules and uses the same basic intelligence that x time produces y heat. Definitely a much better approach for the initial heat up. But then I expect it must enter the same closed loop test and adjust cycles?
That is true. I received this question again and I take the time to respond to it now for all the other people wondering about the same thing. After the ML AI Preheat phase is done, the soldering station enters normal functionality and adjusts the temperature using the normal closed feedback loop and PWM. Then during this normal functionality it samples new heating times for the ML stack.
If a control algorithm just automatically adjusts a parameter to achieve a goal like has been done for 50 years is that machine learning ?
A PID controller parameters can be auto-tuned based on the sound mathematical means.
Machine learning is great when the rules aren’t known, but does a sub-optimal job during the learning.phase. It is like hiring a new grad vs an experienced person for a job.
I use a PID controller on my 55 gallon drum smoker which monitors the temperature and actuates a relay which turns on a small blower fan that feeds oxygen to the coals/wood in the fire basket. It takes a bit to auto-tune like you mentioned. At first it overshoots and undershoots but each time it gets closer than closer and eventually it keeps the temp perfectly at 220’f. mmm… pulled pork. dammit. now im hungry.
Throwing ML on a problem like this is an interesting choice. Why not just throw science and math at it? Calculate thermal mass, electrical heating power and potential energy losses due to heat radiation and convection and calculate how much energy the soldering station would need to get the iron near a specific temperature goal?
Because if you just throw science and math at it, it will be a soulless tool of darkness, and you won’t have an uncorrupted force bond with it, and it might give you an unwanted predilection for wearing black helmets in public. Now, go back and read the title again.
I suppose the advantage here is you can use the algorithm on any iron, in any environment, as the ML should be able to adjust dynamically, without know any of the parameters you brought up.
Is your background CS or mechanical engineering? There is your answer. As an ME, who calculates convectoin coefficients semi-regularly there is a lot to go wrong in heat transfer calcs and plenty of assumptions…
…the very long evolutionary way of “how to not achieve a goal” to sort out the few working solutions….. well, it should not be the engineers path (or the jedis ). Seems more like brute forcing the answer instead of trying to understandingly encounter the given problematic…
I don’t believe in this. for an isolated soldering iron, ok, but as soon as you press it against a place to melt solder, the whole thermal thing changes. no way this setup is going to adapt to constant changing thermal conditions in regular use. Besides, every type of tip has a different thermal constant during isolated heat up and normal use. wrong tool for the job. sorry.
Or do i miss something?
using a 1bit COMPARATOR anyone? thermal mass is low -> low overshoot ANYWAY
Why bother polishing a turd in the first place? Run the heater and feedback out individually. You spend more time and bother trying to fix a bad hardware solution with software instead of just replacing the bad hardware with more logical set up and considerably simpler software..
except the best cartridges work this way, JBC 245 and Hakko T12/15 for example. You will be hard pressed to find better station than JBC for under $350. You will have to go >$500 and look at RF stuff like Hakko FX-100 or Metcal MX-500, and even then you lose flexibility (temperature changes).
The only old style heater inside tip station that still could make some sense nowadays is ERSA I-CON. Huge power and thermal mass, cheap tips, but feedback lags like hell due to distance between end of tip and thermocouple :( and the price is ridiculous compared to modern designs.
Then there is alarm clock UI Hakko FX-951, not worth thinking about. Even $15 chinese ebay T12 controllers makes more user friendly station for Hakkos excellent T12 tips.
I still don’t understand why more people aren’t experimenting with temperature controlled vape mods for a similar purpose. I have experimented a few times with using iron tips (and resistance soldering tips) I’ve built onto RDAs but I used them mostly on unregulated mods because regulated and TC mod firmwares usually have a 10 second firing limit to prevent a lithium ND in your pants. With certain mods however, you can use Arctic Fox or similar modified firmware to control just about everything you would need and then some.