John Bump
5,700 followers -

5,700 followers
Communities and Collections
View all
Posts
Work project: we designed an IC that calibrates itself when it starts up, to reduce the standard deviation resulting from manufacturing processes. It works very well. A result is that because it's a digital calibration, the result is quantized, and under some circumstances, the chip will randomly choose which of two adjacent calibrations it will adopt, so an operational parameter will show a bimodal distribution. The distribution is quite tight: 1% variation across a wide number of chips across a very large number of calibrations. (I've run this 200,000 times on one chip, in one case.) The bimodal distribution is comparatively widespread: 30% difference, with the other peak showing the same narrow distribution. a posteriori analysis of the data set easily distinguishes between the two distributions.
The part that's giving me a headache: the distribution mean shifts a fair amount over temperature. Over externally forced temperature, the two peaks move upwards, maintaining the same distance between them, as the temperature drops. Over internal temperatures as the chip self-heats under heavy load, the two peaks move apart from each other.
I need to bin the results, figuring out whether a particular result belongs to the lower or higher calibration, on the fly, not once the dataset is collected, because I need to write something that allows me to algorithmically find the temperature at which the chip is biased such that it has a 50% chance of calibrating either way, for any given chip (each of which has a different bias temperature.)
I've read some stuff about finding probability density functions (like kernel density estimation) but there's a big gap between descriptions and implementations.
What I'm doing right now is taking a measurement, arbitrarily assigning it to a bin, then taking a running average of subsequent measurements, and taking any new measurement that doesn't fall within a few percent of that average and sticking it in another bin, for which I then start taking a running average, and using the difference between those two averages to make subsequent binning decisions.
The problem with this is that while it looks like it should work, and when I try it out with small datasets it appears to work, and when I try it out on actual data it appears to work, I don't know that it is working and I can't show that it's working without just going ahead and doing the work by hand, which is exactly what I don't have the time to do.
So: any ideas about practical ways to implement near-real-time distribution analysis? Or about how my algorithm could go subtlely awry?
Post has shared content
I've had a bunch of conversations lately about picking and using CAD tools for designing 3d printed objects. makes this look really easy, but this is a neat demo of going from idea to print quickly.
Quick project: Learn how to design this complex-looking tablet stand with simple methods and free tools (Onshape)!
Post has shared content
Post has shared content
When robots act like this, I don't think the uncanny valley is actually all that uncanny at all. It's filled with sympathy.
I have a mouse problem. The house has always had mice, and lots of them: in the kitchen refurb, I took apart part of the floor and found mouse skeletons so old they were just dust outlines of bones. I've tried, unsuccessfully, snap-type traps (including multiple ones beside each other on the theory that a mouse will jump one and land in the next) and seen them sit for two years untouched, I've tried many types of live traps and zap traps. Poison, aka D-Con, works insofar as a box empties within about four months, but I don't like using poison. What has worked well are glue traps. (Well, and my previous dog, who did great mouse-catching work. Current one may be too slow and clumsy.) The ish with glue traps is if you find the mouse quickly, it can be sort of humane. If you don't, starving is an awful way to die. Since I have a lot of traps, many in places that are hard to access, I don't manage to check all of them twice a day. I'm looking for suggestions for how to automate checking/detecting a capture. Small PIR sensor? Sonar/laser? Webcam? Webcam hooked to something that diffs consecutive pictures so it lets me know when there's been a change? Inexpensive but sensitive weight sensor? I would prefer computationally lightweight solutions because I don't want to have 10 RPi's stuck all over the house: arduinos with network capability are way cheaper (and I have them.) Thoughts?
Datapoint: 3M 2" wide blue painter's tape, right on the aluminum bed, PLA, extruder at 230 and bed at 50 with speed at 0.5 for the first layer, then extruder at 220, bed off, and speed 1.0 for the rest of the print, and it adhered fine and pulled off with just moderate finger pressure. This is my first print since demolishing the original tape covering the printer had when shipped, which had adhesion so high I had to pry prints loose with a metal scraper.
Post has attachment
Kishōtenketsu:
http://stilleatingoranges.tumblr.com/post/25153960313/the-significance-of-plot-without-conflict
irritatingly, the included graphic is one cited as a plot with conflict.
Post has shared content
Totally trying this tomorrow.
Flash! "Handheld thermite

Chemistry and materials science teacher Edmund Escudero from Summit Country Day School in Cincinnati has a different take on the classic thermite demonstration. In the reaction, iron oxide (aka rust) and aluminum metal react energetically to yield iron metal, aluminum oxide, and tons of light and heat. Normally, the person running the demo mixes fine powders of the two substances and ignites them using a magnesium fuse, a sparkler, or potassium permanganate and glycerine. But Escudero (who is wearing safety glasses in the photo) simply takes two rusty iron spheres, wraps one in aluminum foil, and then smacks them together. “The impact generates sufficient energy to initiate a small, but dramatic, exothermic reaction,” he says.

Submitted by Edmund Escudero

Do science. Take pictures. Win money. Enter our photo contest here [ http://cen.chempics.org/ ]."
Post has shared content
Artist builds gentle giants from scrap wood and hides them throughout his city

http://flip.it/kZwGnt

Post has attachment
Dropping off bike parts at a local shop, saw this collection of bumperstickers on one of the employees' cars.