Satellite Loaded With AI For Self-Diagnosis

TheReckoning writes "NASA has loaded its E0-1 Satellite with Artificial Intelligence to diagnose on-board failures. The software 'works by comparing a computerized model of how the spacecraft's systems and software should perform against actual performance. If the spacecraft's behavior differs from the model, then the ... "reasoner" looks for the root cause of this difference and gives flight controllers several suggestions of what might have gone wrong.' Another NASA probe loaded with AI was Deep Space 1."

Comments

AI wasted on a satellite

[danwiz]

It just seems to me that it would be better to install more sensors, data-gathering, and reporting capabilities and then leave the trouble-shooting to the people on the ground. Payload costs are expensive, so why put the diagnostic end in orbit?

Given the same data and placed groundside, it could then it could be tuned and upgraded more easily.

I recall how the Mars lander had problems and the ground team worked out a novel solution. I'll bet that they would have like to had extra information to work with, instead of an onboard AI.

BTW - I can understand this approach better for a long-range craft, just not an orbital satellite.

Re:AI wasted on a satellite

[discontinuity]

I would guess that this particular instance is just one step on the path to self treatment of faults. Naturally, accurate diagnosis is a first step in effective treatment.

As for the value of this particular system, someone may have decided that the data needed to run the simulation may take up too much bandwidth to transmit it back to Earth. Don't forget about blackout periods for craft orbiting the moon, Mars, etc and the long lags in comms. This increases the value of having it on board.

-R

That qualifies as "AI"?

[aicrules]

Sure, any automated response can be called AI, but this doesn't impress me.

If all an AI module can do is make objective suggestions, it's nothing more than a list of conditional statements. Whoopideedoo!

I can run similar "AI" on my TI-85. And I could write it all from scratch in the time it takes for a launch vehicle to reach the stratosphere.

The web servers of 10 years ago could "suggest" that an "Object may have moved", so is that artificial intelligence? I guess it's really, really dumb AI....

Re:That qualifies as "AI"?

[Tablizer]

If all an AI module can do is make objective suggestions, it's nothing more than a list of conditional statements. Whoopideedoo!

During the "AI bubble" of the 80's, somebody complained that "any product with IF statements these days is claiming to be AI" (paraphrased). The definition of AI still has no consensus. Who knows, maybe human thinking could be modeled via a database full of IF statements.

Re:So this may be a simple question but...

[pz]

what happens if the AI malfunctions? then mission control will get a bunch of useless error reports...

[sarcasm] Yep, I'm positive that the hundreds if not thousands of PhD-level man hours that went into this part of the project didn't consider that. Yep, took that young whippersnapper Quasar1999 to think about it for a few mintues to evaluate and assess the entire effort and proclaim, "it's a stupid self diagnosis test." [/sarcasm]

If one actually reads the referenced article, it sounds like LV2 is, in fact, something far more advanced than a "stupid self diagnosis test." Se.f-diagnosis tests are pretty straightforward and highly tuned to a specific architecture. I've written something like that to evaluate an experimental compiler, with statements like,

define a=1;
if (a+a eq 2) then print 'simple addition works'

But LV2 is very differnt than that. Into LV2 (which, despite the hype in the article, does not need to be on-board) is built a generic model of satelite functionality customized to the particular device in question. When unexpected results are found, the diagnostic software can experiment on the model, asking questions like, "if, in the model, valve G34 is stuck open, does the model behavior match the current anomalous condition?" I'm sure it wouldn't be hard to write up a test script that could iteratively simulate a fault in one or more parts of the system until it found a handful of likely candidates. Given that there are thousands of components in a satelite, this surely can be done faster by a machine than by a human. Then, were we really trying to do something advanced, we might come up with a way of caching these results to guide future diagnoses and build up a set of experiences. Collect these experiences from different projects (since, if LV2 and its descendent software is widely adopted, the data are presumably in common form), and you can guide designs of future satelites to avoid common failure modes, or identify problematic components.

Now, is that AI? Does it think? You probably wouldn't say so. Could it be an aid to ground-based support? You betcha. Is there a reason to disrespect the fine engineers at NASA by demeaning their efforts without giving fair due? I fail to see one.