AlexGS wrote:In my tests before I saw lot's of noise introduced because of data jamming in Artisan. It was either losing readouts or missing the timing. It resulted in ROR spikes that didn't actually exist. In 1.6.2 this got smoothed out. Plus, for Phidgets, now you can select async at 32ms poll rate. I use it in combination with 0.1C change step and get unusually smooth results with no smoothing on ROR lines and maybe 2 for the lines, but with delta span at 1 sec!!!

Of course, more testing is required.

I never really with the settings in the Phidget tab. Now I am tempted!

Here is what I did to get the result I'm talking about. I used 32ms for faster response and 0.1C to limit number of updates. At 0.3C change the lines get quantized and this makes ROR jumpy again.



The sampling rate is 1 sec.

Marko, Great article and work! This motivated me to look into why my RoR was rather unstable. I like your technique with Idle Noise plots.
What I found was something you may find interesting.

For starters, I am using an older Phidgets 1048_0 (2015 vintage), Artisan 1.5.0 and Windows 7.

Below is a idle plot with 1s sampling interval, no curve smoothing, only 6 RoR smoothing, delta span 1s, no sync, no oversampling and my North roaster is cold with some heating due to morning sun. Notice the RoR is really quite smooth. But at 2min, 5min and 10min I open the door and touch the BT probe for a second. It warms the tip slightly sending the RoR skyrocketing as you would expect. But then as the temps slowly return to steady state, there are spikes in my RoR.



The BT temps looked very smooth. So I zoomed in to temp curves.



Notice the cause of the RoR spikes looks like quantization levels or resolution of the Phidgets. These temp steps are 0.4deg F, which are much bigger than what a 1048 is suppose to have. Could my 1048 be less resolution??

At any rate, others may have this issue as well.

Eddy,

It looks like in you case the signal is quantized at 1F step. I just checked current 1048 specs and it supports 25 updates a second (40ms) and resolution of 0,05C. It should not quantize that coarsely.

Here is an example of my issue in 1.6.0 (I used one of bintray builds for early REL1100 support):


Notice these icicles at the bottom of the curve and jaggedness of the ET in background. The big temp/ROR swings is from TC detecting my finger temp and then cooling down to the air temp.

Here is what the same test gave me in 1.6.2:


Notice that with significantly less of smoothing the curves are much more stable. Both experiments are at delta of 1 sec.

AlexGS wrote:It looks like in you case the signal is quantized at 1F step. I just checked current 1048 specs and it supports 25 updates a second (40ms) and resolution of 0,05C. It should not quantize that coarsely.

Correct, but I measured it at 0.4F steps.
I just searched Phidgets and the old 1048_0 has the same specs with regards to resolution.
https://www.phidgets.com/?&prodid=38

Your change to 1.6.2 looks interesting. The fuzziness on your 1.6.0 looks a lot like my issue! Now I am wondering if I should update to 1.6.2.

On Artisan 1.6.2 I tried Async/0.1C/32ms with no smoothing and 1s DS and got this:



It was very smooth before I introduce heat. But one there was a differential to measure, it was all over the place.

Almico wrote:On Artisan 1.6.2 I tried Async/0.1C/32ms with no smoothing and 1s DS and got this:

It was very smooth before I introduce heat. But one there was a differential to measure, it was all over the place.

This is most likely the effect of this 0.1C Change Trigger. Without heater there is hardly any change above 0.1C thus you get a totally noise free signal. Put that one to 0C and compare.

Here you can read about the Change Trigger parameter. Artisan is just forwarding this setting via the Phidget API to the driver. I would keep the Change Trigger at 0C for most cases.

I need to clarify that I added 0.1C change step just to limit amount of data/samples posted as I was sure that Artisan gets overwhelmed. That is the reason in a screenshot above I had lines smoothing set at 2 -- to smooth the quantization steps. In my testing on my device this combination gave me best result. Though it was extreme I never planned to use for roasting. I'd prefer to have ROR step at 10 sec, which limits amount of data the curve shows and is still fast enough.

I did some roasts with 0.0C quantization step and ROR delta at 10 sec (5 was a bit too detailed but not noisy) and I love the result: on 2mm TC it was registering quick enormous BT ROR spikes I didn't see before (real data, not noise, happened multiple times at the same moment on two different beans approaching FC, investigating).

MaKoMo wrote:This is most likely the effect of this 0.1C Change Trigger. Without heater there is hardly any change above 0.1C thus you get a totally noise free signal. Put that one to 0C and compare.

Here you can read about the Change Trigger parameter. Artisan is just forwarding this setting via the Phidget API to the driver. I would keep the Change Trigger at 0C for most cases.

Interesting. I put it to 0.0C and did a quick 1# roast. Considering DS was at 1s, this is actually pretty good. I had Delta smoothing on 3.



With the new RTD set up this should get even better.

The really interesting thing is how fast the readings are. Look at ∆ET curve and how fast it drops when I make a small heat change. I made a small change at 8:20 as well but did notate it.

Almico wrote:With the new RTD set up this should get even better.

I'd be careful with that. I can not argue with the better results Marko is obtaining. And the internet searches seem to generally say RTDs are more immune to noise. However, the fact is, there are many ways a circuit can measure the RTD resistance. In the case of the Phidget 1200, specs say max RTD current is 62uA. That equates to only 23uV per degree C at 200C. Compare this with a K thermocouple, which is 41uV per degree C. And the thermocouple has much lower resistance. So in theory, the noise immunity of the thermocouple should be better. But that depends on other factors such as common mode rejection, which is not specified.

The analog filtering, which is very desired, is accomplished in the tip of the RTD. The noise likely couples to the wires after the RTD and so this anolog filtering does no good IMO. The best place for filtering is at the A/D converter within the Phidget. Putting a capacitor at the terminals would do exactly this.

That said, a better RTD circuit would put more current through. I've read 1mA is pretty common. But higher current results in some self heating. So there is a trade with accuracy.

Lastly, I would be a little worried of ruggedness of cheap RTDs. Often you see that they cannot be bent. Buy commercial probes would likely mitigate this issue.

But RTDs, IMO can be used to achieve very robust, highly accurate and noise immunity better than thermocouples.
I'd be inclined to construct a circuit using the Phidget voltage sensors.