2 ms·

> I wouldn't say this is actually true: while there are generally fewer dimensions in hardware testing, it's not actually a given that if you have tested over t

> I wouldn't say this is actually true: while there are generally fewer dimensions in hardware testing, it's not actually a given that if you have tested over the range in each of them that you will certainly be OK across the entire space I agree that the point where the power supply dissipates the most power, or the point where it starts making an audible whine, might be in the middle of its operating range rather than at its limits. You would certainly want an experienced engineer who can say what density of parameter sweep is appropriate. This is what I was trying to say when I said that when testing a system designed to operate between -5°C and 55°C, expert knowledge might assure you there was no need to test at 21°C if it had been tested at 20°C and 22°C successfully.

nunuvit 3y ago
I've designed power supplies and I can tell you that the "density" of the test doesn't matter so much as the validity of the test. One way we assess stability is by doing frequency sweeps. What allows us to infer stability from that test is the assumption that the system is linear time invariant (LTI). We assume it's LTI because we tried to design it to mostly act that way, even though it's really not. For an LTI system, a frequency sweep and a step response are completely equivalent and interchangeable, it doesn't matter which one you use. But what actually happens is the step response test reveals different information. This can only happen if the system is not LTI. Therefore neither test is conclusive about stability, though they're still informative, which is why we do them. 10x density would be no less inclusive. Another problem is that some stability issues have very low observability. The blip or offset they cause in testing is indistinguishable from normal switching noise and measurement imperfections. They only really reveal themselves when they become a problem in the field. 10x density wouldn't find the issue because it's caused by some other combination. Explaining and fixing those issues is tough, and anticipating them requires an arsenal of models developed from someone's hard earned experience. I don't mean to diminish the density issue. That's still a real thing, but it's kind of an orthogonal problem. The formal term is optimal design of experiments, and in simulation we use search methods like Monte Carlo because it's not feasible to enumerate the design space even on a computer, much less in a real test.

nunuvit 3y ago

I've designed power supplies and I can tell you that the "density" of the test doesn't matter so much as the validity of the test. One way we assess stability is by doing frequency sweeps. What allows us to infer stability from that test is the assumption that the system is linear time invariant (LTI). We assume it's LTI because we tried to design it to mostly act that way, even though it's really not. For an LTI system, a frequency sweep and a step response are completely equivalent and interchangeable, it doesn't matter which one you use. But what actually happens is the step response test reveals different information. This can only happen if the system is not LTI. Therefore neither test is conclusive about stability, though they're still informative, which is why we do them. 10x density would be no less inclusive. Another problem is that some stability issues have very low observability. The blip or offset they cause in testing is indistinguishable from normal switching noise and measurement imperfections. They only really reveal themselves when they become a problem in the field. 10x density wouldn't find the issue because it's caused by some other combination. Explaining and fixing those issues is tough, and anticipating them requires an arsenal of models developed from someone's hard earned experience. I don't mean to diminish the density issue. That's still a real thing, but it's kind of an orthogonal problem. The formal term is optimal design of experiments, and in simulation we use search methods like Monte Carlo because it's not feasible to enumerate the design space even on a computer, much less in a real test.