If Sense cannot identify a device after looking at it and watching it start THOUSANDS of times it’s certainly not going to suddenly be able to identify that single spike as a “bad cap” failure.
At best it might be able to say…“Whoah…this is not something we’ve seen before…perhaps you should look at it”.This exactly what we can do today by manually adding the peak goal.
Imagine a failure signature, for arguments sake, is a power spike of a certain current range lasting for a certain length of time and is otherwise “flat” (e.g boom 100A for 2 seconds) … Sense is more than capable of detecting that regardless of device detection. If that weren’t the case, the Goals and alerts wouldn’t be working.
Of course, a true failure detection is going to present a more complex signature … but it has little to do with the separate device detection and is arguably a simpler pattern to search for, especially in the extreme current case that is typical for a failing HVAC.
Venstar makes sensors that are wifi battery powered and communicate thru their Colortouch thermostats. One can be installed on the supply leaving the AC system and one on the return duct coming into the system. you can then set up alerts for supply air being too warm. It’s a simple way to see a problem developing.
Heating and air conditioning systems only deliver an average of 56% of the energy they consume to cool the space. Your system can be tested by any air conditioning technician certified by National Comfort Institute to find out what percentage of the total Btu’s your system gets into the conditioned space and how to correct it.
Most bolt-on monitoring systems focus on the equipment and refrigerant circuit. They ignore airflow and the ducting which is where most of the energy the system uses is lost. Testing my AC unit and correcting the ducting lowered my utility bills 27%.
I have thought of ways to address similar concepts. I would add an overlay on the power consumption graph indicating the outdoor temperature for the zip code where the sense monitor is located. you can then get an average runtime normalized for outdoor temperature for their particular air conditioning system and area. then if the runtime begins to increase beyond a threshold it alerts the owner about the possibility of a loss in cooling capacity.
@lee, @bright1135 's recent comments reminded me that you had posted this question last year. About a month ago I installed Ecobee temp sensors in my upstairs and downstairs registers closest to my furnaces. I’m using those sensors to monitor performance. I have been using the sensors to chart weekly register temperatures when my AC is off and on. I’m hoping I’ll be able to detect declining cooling performance, though the cooling season is coming to an end. Looks like I might get one more shot with a hot spell this coming weekend. I think any AC system that includes remote temperature sensors should include this kind of monitoring.
Where I work, we call this the Predictive Maintenance problem. Monitoring signals and detecting conditions of interest, especially ones that portend trouble. Maybe googling that buzz phrase will be helpful.
Do you mind me asking what you do, @don.mathis? It seems you’re pretty knowledgeable about HVAC / home efficiency in general, and it’s always great to have another resource to bounce questions off of 
Thanks but I actually don’t have any HVAC knowledge! I’m a software developer at The MathWorks (the people who make MATLAB). I’m part of the team that develops the Statistics and Machine Learning Toolbox. One of the products built by another team is called the Predictive Maintenance Toolbox, but I don’t know much about it, just the general idea. In any case I’m definitely not here to plug products. But feel free to bounce machine learning questions off me! I saw something about data download on another page here, maybe some day I’ll play around with some of this data I’m getting.
@JustinAtSense,
This is more in the Ecobee camp, but I bet they could train on thermostat system mode, ambient temperature for the HVAC air handler / evaporator coil, input temperature (temperature at thermostat), and some form of system runtime (how long has system been in that system mode) vs. output register temperature, to predict future register temperature. Then if the output register temperature deviated very far from the predicted temperature, one would know there is an issue.
I have another partial heat spell worth of data for my tracking. Things are looking reasonably good, though there is a big difference in output temperature performance between my upstairs and downstairs units. A few possible reasons:
Trying to look at one more measure of performance, time response of register temperature vs AC runtime. Wondering if we have any thermodynamics experts out there who can help me out ? What should the analytic solution look like given a starting temperature T0, an intake temperature TA and a ramping up cooling system ?
Here’s register temperature vs. runtime for the start of a bunch of cooling cycles. For an EE like me, that looks a lot like an RC relaxation curve.
But if I look at the ln(register temperature) vs. runtime, it’s pretty clear this doesn’t fit the relaxation curve of the form T0 x e^^ -t/RC type of form. If it was, we would see straight-ish lines.
Maybe it’s a power curve, and I should look on a log / log scale - better !
Or is an inverse curve ? That seems to give a linear region that I can pull characteristics off of.
Any thoughts ??
Very cool indeed! Newton’s law of cooling may be helpful here.
I think you have a bunch of different interactions layered on top of each other:
-First, are we assuming a fixed blower speed in the indoor AHU? If not, the ramp profile of that blower will affect the delta T across the coil.
-The indoor coil heat removal rate depends on refrigerant flow rate. At startup, getting pressure drop across the metering device, and therefore flow, requires the condenser pressure (and therefore temperature) to rise, which takes time as the coils heat up. I’m not entirely sure, but I bet this is a logarithmic relationship to time.
-There’s a bunch of thermal mass in the system (the coils themselves, refrigerant piping, ductwork, etc.) that needs to change temperature. This is newton’s law of cooling, referenced above, but we’re not interested in the temperature of the objects that are cooling down, but instead the heat rate going into the airstream. I think this will again be a logarithmic relationship as the delta T beween the heat-losing objects and the airstream decreases.
So, I think you have a few logarithms stacked on top of each other here. Modeling this accurately would be pretty complex.
@lee, thanks for the pointer. I knew the solution to the simple thermal equation was of the same form as an electronic RC network - a T0 x e^ -kt /RC kind of equation. But your link also reminded me of two things.
@pswired, I’m guessing you are right. There really are two or even three different domains to this problem.
ps: I have a constant speed air handler and single stage compressor.
Here’s one more view of my downstairs and upstairs HVAC performance looking at the Delta between the register temperature and the overall lowest measured register temperature over the course of all recent AC runs greater than 50 minutes. If I was really looking at the right delta I would look at the register temp vs the temp right at the evaporator coil, but I have no way of measuring that realtime right now. These curves give some idea of an envelope of OK performance. Now I just need to find a way to characterize the envelope for each !
I’m still poking around at the mysteries of my HVAC cooling as the cooling season tapers off to see if there is a reliable way to predict a deterioration in cooling over time (performance monitoring). Here I’m looking at things slightly differently - looking at the first 2000 seconds (33.33 min) of all continuous cooling cycles 30 min or greater. The Y axis is the Delta between the current register temperature and the minimum register temperature over the entire cooling cycle. One of my goals was to dial in on the initial cooling regime where the cooling might be following a simple Newtonian heat law. I also removed the legend, to focus in on the characteristic curves.
Looking at the downstairs cycles above, it looks like there is a 250-300 sec delay before the evaporator coil begins to cool the air in the nearby output register. That’s probably attributable to the startup of the compressor and flow of the coolant to the evaporator coil. The upstairs cycles indicate a longer delay, closer to 400-500 seconds. But that makes sense since the coolant has to travel roughly double the distance from the compressor. The upstairs evaporator coil is in my attic so it also experiences much hotter ambient temperatures most of the time when the AC is cooling.
I’m thinking that there might be a way to peel away the “ramp” period to get to pure cool down curves that I can then fit to an exponential function.
