Genetic Algorithms

Wade Wrote:
-------------------------------------------------------

> I wonder, does anyone make a sort of instrumented
> bolt for a reasonable price? The extruder mount
> could use a pair of bolts with strain gauges, and
> you could watch the change in bolt tension when
> you run into something. Probably cost as much as
> the RepRap.


Wade, et al,

If we had such a sensor (crash/no-crash), how would it be used? Stop everything and wait for help? Or is some more complicated, automatic recovery envisioned?

Generally, strain gauges (particularly applying them) is labor intensive (read fussy work) and in consequence usually not cheap. The signals from foil strain gauges are small, so one needs high amplification (with low time & thermal drift) to get a usable signal.

Metals have much lower failure strains (order of 1e-3) than plastics (10 -- 1000x as a WAG), so instrumenting a plastic "bolt" or other structure would be easier -- except that plastics are terrible heat conductors, and have much higher thermal coefficients of expansion, which makes it difficult to match coefficients, such that one can measure mechanically-induced strain, separately.
Plastics generally creep, if you strain them long and hard. And some are virtually impossible to glue strain gauges (or anything else!) to them.

However, there may be a way to skin this cat.... I did some preliminary experiments (read, played around with) some conductive-plastic fiber I got a sample of, and it serves as a pretty good strain gage, and well matched to plastic, because it is plastic. Since it has high intrinsic resistance (compared to foil gauges), it develops a larger signal, with minimal resistive heating.
(The mfgr. didn't seem very interested in this aspect -- may have been too far out of their area of expertise....)

Typically, bolts aren't the best shape for instrumenting, even if one turns down the thread for a flat surface (and smaller section to increase the strain) because bolts are usually strained mainly in tension. Load-cell structures are typically designed to strain the instrumented areas in bending, torsion (or sometimes shear), to get us usably large strains. I don't have a darwin, and I'm a bit hazy on the attachment of the extruder to the cartBot. If you (or anybody else) can post some photos (or sketch w/dimensions) of the attachment, I might be able to come up with a mod that'd serve as a load-cell structure and try putting some of my "strain gauges" on it.

I know we'd *like* three forces and three moments, resolved about a common point, e.g. the extruder tip. However, a 6-DOF load cell is usually a complicated (hard to machine) structure, and usually needs 6 -- 8 pairs of gauges.
Can folks with extruding experience suggest what the minimum set of forces/torques would be? And what part of the reprap usually hits things, the tip, or someplace else?

---

Larry Pfeffer,

My blog about building repstrap Cerberus:
[repstrap-cerberus.blogspot.com]

Here's a force sensor that should be fairly easy for use to use in our environment.

[www.hvwtech.com]

@Forrest: A high data rate between reprap and PC shouldn't be necessary. What I was envisioning was more of an auto-calibration process done as a pre-process before printing. Basically, the rerap would print a simple test pattern, use the imaging system to measure the accuracy of the pattern, modify it's print parameters based on the image information, and then repeat the process until calibrated. This would save the need for someone to hand-adjust the each reprap/repstrap for temperature, flow rate, and slop/backlash in the support frame/stepper motors. If done well, it could even prove to be more efficient at it than all but the best human operators (similar to how ABS breaks, in cars, can react better than is possible with human reflexes). This strikes me as an important feature to have in Reprap if the project is to ever allow for mass distribution as most people you are looking to distribute to won't have a fraction of the mechanical aptitude you guys have.

As for the 3d scanning feature, that might be slower due to direct control of the print head from the computer but shouldn't be all that time consuming based on the DIY 3d scanning algorithms I've seen posted online.

@Wade: True, recreating a computer vision library from scratch would be difficult, but there might be some open source resources that might speed the process up such as the open source OpenCV library released by Intel and a similar library released by Nvidia (the name, of which, escapes me at the moment).

The primary advantage I see in a webcam, as opposed to accelerometers or strain gauges, is price. Cheap webcams can be had for as low as ~$10 (though I don't necessarily suggest going for the absolute cheapest ones). The resolution and color quality won't be the best, but moving the print head closer to the target can get around resolution and color quality doesn't matter as much since most computer vision calculation are done in black and white anyway.

-Shawn

Wade Wrote:
-------------------------------------------------------

>
> I wonder, does anyone make a sort of instrumented
> bolt for a reasonable price? The extruder mount
> could use a pair of bolts with strain gauges, and
> you could watch the change in bolt tension when
> you run into something. Probably cost as much as
> the RepRap.

If all that's wanted is a (binary) crash sensor, then a simpler approach (than strain gauges or accelerometers) might be a springs & contacts approach, similar to what's commonly used (in a steering wheel) to sound the horn, no mater where you press on it.

---

Larry Pfeffer,

My blog about building repstrap Cerberus:
[repstrap-cerberus.blogspot.com]

I'm more interested in closing the loop on extruder height control - I'd like to know exactly how high my extruder head is above the bed, as height changes of 0.1 mm make significant differences in raft and part adhesion. My current method is on the crude side. The PTFE length changes with temperature, and time, and sometimes the frame itself warps over time, so I'd like to measure the important variable - nozzle to bed height - directly.

Collision detection is a nice side affect; I'd have the machine shut down and flag me if it happens, as opposed to breaking something. Right now I have the Arduino shut everything off, print an error message to the serial port and go into an infinite while loop when something really bad happens, like a thermocouple malfunction. That saved my machine once already.

That IESP-12 force sensor is priced right at $7, but I'm a little skittish about the "estimate" language in the datasheet. It doesn't sound like it works well over long periods of time under force though, which could be bad in our setup if it's always measuring the weight of the extruder.

A metal bar with some strain gauges could work too, but the heat flow from the extruder might mess with things.

In general, this sounds like a little more fun that I'm interested in at the moment. And no, this has nothing to do with genetic algorithms, sorry.

Wade

Edited 1 time(s). Last edit at 01/15/2009 10:38PM by Wade Bortz.

Wade Wrote:
-------------------------------------------------------
>
> That IESP-12 force sensor is priced right at $7,
> but I'm a little skittish about the "estimate"
> language in the datasheet. It doesn't sound like
> it works well over long periods of time under
> force though, which could be bad in our setup if
> it's always measuring the weight of the extruder.
>
> Wade

Hi Wade

My idea was that it would measure upwards pressure from a calibration touch down or abort if it had run into the bed.

It could also be used for a calibration cycle where you touch down on the four corners of the bed or the maximum build area to check for alignment or height of Z; which in IMHO is the measurement that can be the most variable.

The other idea that I had was measuring contact force for a stencil cutting blade or drawing pen to get more uses out of the bot.

As a stencil cutter it could cut logos, labels, paper or PCB exposure or silk screen masks; other capabilitys to show off why we are into this 3D stuff.

Since it is variable resistance it could interface to one of the analog ports for the Sanguino in place of the second extruder which no one has really done yet.

Though if you still wanted to have a second extruder you could do away with the separate Z-enable (tie it to X & Y enable for a generic stepper motor/movement enable) and then devote that pin to Z contact pressure.

You know that

Edited 3 time(s). Last edit at 01/15/2009 11:48PM by freds.

The upward pressure is the same thing general idea, but if you've got it truly not touching the sensor (as that datasheet reccomends), then you've got a certain amount of backlash to take up before the sensor registers, which would prevent me from knowing exactly when the head touches, which is what I'm really interested in. There's fair amount of give in the entire system when you're looking at scales of less than a mm, so I was thinking a constant measurement of the force on the head (including it's weight) would be useful. That points more towards strain gauges.

Yeah, the 4 corners calibration was what I was thinking as well. It could also pick up on a tangled plastic filament feed, which has wrecked a few builds of mine in the past as well. That would probably just trigger the head collision detection, due to the upward force on the extruder.

Mostly I just squint at the webcam feed every few hours and try and deduce if something bad has happened.

Wade

A cheap bump sensor is described on instructables at:
[www.instructables.com]

Basically, stick a metal rod through a pipette extending out a bit, put a spring around the pipette extending also, and check for continuity. For the reprap, the spring & rod would be pointing upward, like a car antenna, and whenever the nozzle bumped into something, as long as the acceleration was high enough, the spring would then bend and bump into the rod. The continuity would trigger the switch circuit:
[www.instructables.com]

and a signal would be sent to the Arduino to shut down.

If the sensor trips in regular operation, the spring would be trimmed bit by bit, until it doesn't quite trigger during operation.

In that bump sensor, the red wire is connected to the middle of the spring, but I think it would be better to connect it at the base of the spring.

What is also possible is just having a long thin rod in the middle of a tube. When the rod touches the tube, the IO goes zero...

Same principal, different layout.

Another option to detecting height might be to use several hall effect sensors embedded at strategic locations in the print table, and a magnet on the print head?

Found some cool resources:

[blog.makezine.com]

[www.cs.nyu.edu]

For those more technically inclined:

[www.cogs.susx.ac.uk]

The Hybrid vigour side of GA is fascinating.
There's an interesting site:
[www.rdos.net]
Which goes over an alternative to the mitochondrial DNA evidence for the African "Eve" theory. Basically, it's saying that the "proof" was more along the lines of "absence of evidence", than "evidence of absence", and that the African emigration occured in waves. As a result, recombination of the partially specieated peoples, (probably the rape, pillage and plunder scenario), resulted in some interesting mixes in following generations. Certainly, the more competent level of inventiveness seems to have flourished from some 20,000 years back, and onward. The site points out various gene "defects" that do not appear as defects when the environment is altered to be more like some of the precursor environments. One of the interesting points made, was the speciation of kept beasts, away from the wild animals, and of the DNA changes over time that point to at least one separation/speciation period, which is far in excess of that proposed by the Out of Africa proponents.

On another site:
[www.evolutionpages.com]
is some interesting data on our critical difference from the other great apes, if I recall correctly, they have 23 chromosome pairs, and we have 22. one of our pairs appears to be a splice of two of theirs. I like to think of us as being like Win CE, the cut down version of windows.

Back on track though, I suspect GA are a little like fuzzy logic. Forest points out that to create the environment in which they would be fruitful, is a huge mission in itself. "Expert systems" are what most programs today can be defined as.

In the blood and bone world, "mother nature" practices the ultimate determinism, trialing solutions in the actual environment. In that scenario, successes are not always rewarded, and failures are not always culled.

In the reprap development environment, the determinism aspect is not so honest, with people watching the experiments, there is bound to be tampering with the results. However, sometimes people get lucky, and notice something that works, nurture it, and in time, we have a new feature, that has escaped a premature cull, for some other reason.

Tools continue to advance, and given another 10-50 years, it's certainly not inconceivable, that AI systems may have environmental sensing systems good enough to take deterministic modeling many significant steps forwards.

The human factor... As our tools advance, the disparity between their and our intelligences diminish, normally, the human race has needed people like Nicola Tesla, Leonardo Da Vinci to lead the way. The Nicola Teslas aren't in every generation though, and so typically we've had, as a species to wait many generations for the next breakthrough/s. As our tools become more A.I like, the human factor in co-evolution is going to become more of a "flavour", than a control, you have only to look at technologies developing around us, to see the progression, and our place within the progression becoming less of masters, and more as symbionts with the machines we create.

Graham Daniel