| I’m not disparaging their work. It is probably really cool, and they probably published some great information that will be useful to many. I don’t doubt it was challenging for them, but I do doubt that the problem was fundamentally challenging from en engineering perspective. As for my “work” it is literally insignificant tinkering by a bored old fucker with nothing better to do than chat on hacker news.. I don’t even respect my work, and anyone who thinks more of it than digging a ditch is just wrong and has obviously never dug a ditch. But, just calling it like it is, the “signature “ thing they are working on is something that is already solved for decades and if it took anyone more than a week they may not have a clue what they are doing. I have implemented a version of it myself in a technically adjacent application. In case anyone cares enough - and you probably shouldn’t- feel free to read my incoherent ranting that follows: In my case I use load vector analysis it to detect and characterise loads on our microgrid. We have several buildings and houses, and we run 100 percent solar on an off grid system. Using an esp32 and a current transformer coil on each of the three phases, with some good 16 bit ADCs, we monitor and characterise loads. Each of the refrigeration compressors has a somewhat unique starting and load profile. Each water pump in our utility system similarly has a unique startup and load profile. Same with air compressors, fans, and other equipment. The profiles are programmed into the esp32 by putting it in calibration mode and switching the load off and on 10 times. It’s a pain in the ass because you have make sure no big changes happen in the power system in the meantime, but it works. The MCU saves the signature as a vector and assigns it a number if it doesn’t sit too close to any existing vector signature. It is really good actually, even being able to discriminate between identical pumps on the system because of their supply impedance and loading. I’m not a data scientist or an actual engineer so I adapted some vector code from a DSP project, and the whole thing took me about 2 days using the Arduino IDE (please kill me) I’m basically an idiot. Anyone who does this for a living should be able to do it in less than half the time. There are still some rare false negatives because a grid can be quite chaotic, but in general it’s very accurate. In a simple D.C. system like a car in the off condition with predictable loads I would fully expect 4 nines discrimination. What they did was cool, but it wasn’t hard. Not saying it wasn’t hard for them, and maybe they learned a lot, but I’m pretty sure that 1.2 million to solve the problems described in the article is two orders of magnitude off of reasonable. From the provided description, If a single engineer with decent tools could not have this from zero to a production ready GERBER file with masks, stencils, and the works to send off for automatic assembly inside of a month they should probably look for another line of work. Of course, if they work like I do which is to say they don’t, very much, and they mostly drink coffee and fuck off all day, then I’d give them a month and a half knowing full well they did all of the actual work in a week of panicked thrashing, creating months of technical debt in every line of code to build the glass house that somehow works without passing any of the tests but that’s fine you just rewrite the tests. Of course certification and things like that are a whole different beast, but this was a CORE research grant. |