Listen to the Machine; The Machine is All-Knowing

So, what happens when the machine is lying to you? That has happened to me before, though in a more benign if irritating examples. Such as a phone application telling me the payment processing system was down, when it was working perfectly.

So I have learned to hate the machines. They lie all the time.

There are plenty of reasons to "hate the machines" without painting them all as pathological (mechalogical??) liars.

Note that my "resolution" is not to BELIEVE the machine. Rather, to *listen* to it -- instead of dismissing its observations without careful consideration.

In the first anecdote, the odds of the ACbrrr actually *failing* (while still apparently blowing cold air) were considerably less than there being a bug in my implementation that was manifesting as this warning.

In the second, I was building a second machine IDENTICAL to a first located nearby. The fact that they were (apparently) behaving differently was likely caused by a difference in BIOS revisions, or a screwup in my building/configuring the disk image.

In both cases, the machines were behaving exactly as designed and it was PEBKAC to blame.

There's big money in AI supported predictive maintenance these days. AI-driven predictive maintenance (PdM) for HVAC controllers relies on a continuous stream of real-time sensor data, historical telemetry, and ambient environmental conditions. Machine learning models analyze these factors against dynamic baselines to spot anomalies weeks before a physical failure occurs.

Thermodynamic & Fluid Parameters

Those require a significant number of new and expensive sensor installations that are not normally included in off-the-shelf systems. Maybe they are now.

Mechanical & Kinematic Parameters

Electrical & Power Parameters

Those are probably the most suited to backfit in existing systems. The sensors required are el-cheapo, and the rest is canned processing. Things like power draw and duty cycle monitoring can be very revealing, especially if they start to radically depart from what the system knows is the norm.

External & Boundary Parameters

It would be nice to see that kind of data, especially if the performance degradtion is graceful.

Just curious, but why are you designing an HVAC controller?

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

Exactly. Your "thermostat" only has one (realistically) input and one output (some "communicating" controllers try to add more).

It has no way of noticing the temperature differential between intake and outlet air. The temperatures of liquid and suction lines. Internal humidity. Outdoor temperature, wind speed and direction, cloud cover, local flora, etc.

Nor can it model the relative effectiveness of evaporative cooling vs. refrigeration *or* predict if using one (or the other)) is likely to be effective on a given day of the year (e.g., if outdoor conditions

Similarly, you can't dynamically alter the routing of "conditioned" (heated/cooled/etc.) air to individual parts of the house without carefully monitoring the back-pressure the plant experiences (some modern units will gag at more than 1" water)

It can't know if the filters need to be cleaned. Or, if the cooler pad has calcified and requires replacement. It's up to "dumb", open-loop constraints ("change the filter every 30 days") or *observant* occupants to determine the level and scheduling of service for these things.

Accessing the actual coils requires more invasive instrumentation -- likely "voiding warranties" (or, at the very least, confounding run-of-the-mill service agents); and, that instrumentation is part of the equipment (which can be replaced) instead of part of the residence: new equipment means new effort to instrument.

I think mini-splits include some of these sensors to keep their efficiency high. But, there's no published/documented/blessed way of accessing that data. And, it most certainly wouldn't be "standardized".

You don't need to instrument the plant -- if you instrument the service (electric and gas). You can infer the portion of the current consumption attributable to a particular appliance, esp if you can tell when it is on vs. off and correlate that with the service observations.

By itself, this isn't sufficient. You need knowledge of the edifice's construction (how much glass surface -- and where; wood vs. masonry, etc.), exposure (sun, wind), thermal mass, etc.

You also want to be able to *anticipate* the occupants' actions and requirements. After all, the edifice has no real "needs" of its own.

The simplistic {wake, away, return, sleep} model of occupancy -- even if it accommodates weekends AND holidays AND illness AND... -- is just too naive. E.g., we "kiss" the furnace in the early hours after rising just to take the chill out of the air (having slept with open windows to deliberately ventilate and cool the house overnight). Yet, switch to air conditioning soon after.

Why should I have to pre-arrange an occupancy schedule -- let it NOTICE when I'm not on the premises. If I drive off at the same time every day and stay away for many hours, it's PROBABLY safe to assume that my driving off at the same time *today* will result in a similar absence.

I I drive off in the middle of watching a movie (that I've explicitly PAUSED), then there may be a chance that I've just run out to attend to something "transient".

If you "notice" me returning early, you can always adjust the heating/cooling cycle to accommodate my unexpected return (without me having to grumble about the house being too cool or too warm and MANUALLY compensating)

few of these things can be done in *an* appliance -- they all require coordination between many such "appliances". And, if produced by different vendors, some standardization as to the data/actuators available and the means of accessing all that.

Smart devices are a nuisance. I still haven't figured out how my Audi decides whether to stop the engine when waiting for a red traffic light. Sometimes it does, sometimes it doesn't. It restarts the engine when the car in front moves by a meter or so, but it's not smart enough to see that the traffic light is still red. Sometimes it restarts when I remove my foot from the brake pedal, sometimes it does not. Sometimes it stops the engine in a down slope while using the speed regulator, sometimes it doesn't. Sometimes it stops and then restarts two seconds later. I think the driver assistance mistakes a roadside beacon for an obstacle, but really I have no idea why. All this interferes with my driving.

There is a good second of delay between my acting on the gas pedal and the car's reaction. You learn to anticipate, but it's still a nuisance.

Sometimes it puts up a message in the dash, several lines of text explaining that "driving assistance" is degraded, usually just at moments my undivided attention is needed most. It's a royal nuisance. Audi got several things seriously wrong.

A really smart device it one that you can forget when it works.

Jeroen Belleman

VCDS (and many other decent OBD diagnostic tools) can turn off the start stop in most VAG cars.

Wanna bet it has very limited information about the vehicle's current "context"?

I.e., you likely have GPS and navigation handy -- yet don't integrate that information with your decisions as to how to "tune" the car's activities. You don't need to "see" the state of the traffic signal you are approaching to know that it is there! And, how "far back" the driver starts slowing gives you an idea of the depth of the queue of waiting vehicles (from which you can guesstimate how long they've been waiting -- how long before the light changes!)

Many navigation systems have speed limits as part of the stored graph -- so, you know the driver isn't LIKELY to be accelerating past 45MPH on this stretch of road (unless you KNOW him -- from continuous surveillance -- to habitually "speed")

I was discussing smart algorithms for application in auto racing -- where the driver only has a PTT switch (so can't "tell" the vehicle's controller what he THINKS would improve his chances.

"Gee, you KNOW what the track looks like after one lap. So,m you KNOW where the straightaways are, how long they are, where the turns are (and which direction), etc. Clearly, the car can ANTICIPATE how you are going to drive given its knowledge of where you are on the track and how you have navigated that stretch the past umpteen times!"

Low on fuel? Maybe we should switch to "economize mode" (you know how many stops you've already made and how long ago -- miles/gallons -- the previous stop) instead of performance mode?

The problem lies in approaching these problems with insufficient information (context) -- which COULD be available had you opted to talk to other (sub)systems that are available.

Knowing is not being right. (S.

It doesn't have to be right. It's comments just have to "have value".

Don Y snipped-for-privacy@foo.invalid wrote: |-------------------------------------------------------------| |"[. . .] | |[. . .] (having slept with open windows to | |deliberately ventilate and cool the house overnight). [. . .]| |[. . .]" | |-------------------------------------------------------------|

A friend likes to sleep with an open window but is this a good idea? Nights here (here is not America) are more humid than e.g. 2p.m. Cf. graphs etc. shown by

Jeroen Belleman snipped-for-privacy@nospam.please wrote: |----------------------------------------------------------------| |"A really smart device it one that you can forget when it works.| | | |Jeroen Belleman" | |----------------------------------------------------------------|

During recent years the general public is impressed by "AI", even more so than when dual-speed CD-ROM drives used to be accompanied by boasts about "multimedia". It is a trite ex-saying by AI professionals that persons who do not profess to work on AI deem AI to not be AI when it is useful. A lecturer uttered "A.I. my arse" during Academic Year

During this week I submitted a form which asks was AI used to answer that form. (S.

Don Y snipped-for-privacy@foo.invalid wrote: |-----------------------------------------------------------------------| |"> Knowing is not being right. | |It doesn't have to be right. It's comments just have to "have value"."| |-----------------------------------------------------------------------|

Indeed. (S.

We have 5 seasons: Spring (Hot), Summer (VeryHot), Monsoon(Hot&Wet), Autumn (Hot), Winter (warm).

With the exception of Monsoon, it is very dry, here. (We have a short "rainy period" in Winter but it doesn't come with the same high humidity that accompanies Monsoon)

A typical day is a 30 degree spread (low to high) so while it may be 85F today, it will be "high 50's" at night. The thermal mass of the building means the structure takes forever to shed the heat it has accumulated throughout the day and the preceding months.

So, to cool the house you have to swap the air with cooler air.

This works great when the air is dry (it happens to be raining for the third consecutive day, here).

Unfortunately, it isn't really possible to automate opening windows -- beyond "roof windows".