Calibration Gas (NO2) sensors

[NeilH]

Looking at the MiCS-2710 base resistance and sensitivity its obvious to me that the Gas sensors need calibrating at manufacturing for the subsequent sensor readings to have any validity. 

The data sheet states a basic fact of sensors right at the end

“e2v semiconductor gas sensors are well suited for leak detection and applications requiring limited accuracy. Their use for absolute gas concentration detection is more complicated because they typically require temperature compensation, calibration, and sometimes as well, humidity compensation. Their base resistance in clean air and their sensitivity can vary overtime depending on the environment they are in. This effect must be taken into account for any application development.”

I haven’t heard any reasonable engineering propositions (algorithms please) for the raw sensors readings.

If I’m wrong and missed it please point me at the right place. 

However since this is a limited accuracy, and basic Air Quality is required they can indicate absolute levels if calibrated. In my opinion the calibration is the least expensive if calibrated in quantity as part of the manufacturing process. 

So, I’m wondering does anybody have any experience of calibrating Gas (NO2) sensors – setup equipment and source supplies. 

It seems like a basic calibration at manufacturing is do-able with a simple test rig – this is the maker culture after all.

From the mics2710/2714 document it seems it would need to provide an enclosed space with an atmosphere of “Synthetic  Air”,

 and then be able to introduce standardized concentrations of the gas NO2 for 100ppb and 250ppb

I’m suggesting 100ppb as the current Air Quality Health standard 

http://airqualityegg.wikispaces.com/Hardware-Sensors     Sti Pollutant Table.doc

http://www.environmentalleader.com/2010/01/26/epa-puts-in-tough-no2-pollution-limits/

and 250ppb as per the Rs/Ro Sensitivity setting in

http://airqualityegg.wikispaces.com/Hardware-Sensors    MICS-2710

The STi “NO2 Sensor Report.pdf” - http://airqualityegg.wikispaces.com/Hardware-Sensors   Fig 2-4 shows a potential ventilation space that could have a sealed test space inside it. 

Seems like that would be possible for the test engineering group to build.

For a manufacturing test, the devices under calibration (DUC) are placed in the sealed test space of known volume with a one way pressure release valve. Of courses there would be many DUC place in it at the same time.

Initially its filled with Synthetic Air @23C – and the baseline reading taken and stored in DUC EEPROM.

Then another known volume of NO2 is passed into to that space to provide a known dilution – ie 100ppb

Does anybody have the dilution volume equations to hand that can speak to this.

I was thinking of a sealed a flexible bag to the side with a simple valve. The bag can be inflated to known volume with NO2, then squeezed into the main chamber with the DUC.

Then all the DUC make a calibration measurement and store it in the DUC EEPROM

Then this process is repeated for 250ppb (or even 200ppb)

Some background of designing a transducer from a raw sensor – the raw sensor is the enabler for measuring a physical quantity. 

The engineering language is that the parts – hw circuits, software & mechanics need to be integration tested to VERIFY that the numbers make sense, across the temperature range of interest.

The process of turning that measurement from an ADC to a number with units eg ppb – takes experience and proving it works.

The theory is established before hand, the individual components of hardware and software and test environment defined, and then its all brought together – with modifications usually to algorithms – to prove the correctness of all the parts.

[David Holstius]

Hi Neil, I have some experience with this, and so do some of my
colleagues who are also studying health effects of air pollution. NO2
is of particular interest to us as a marker of traffic-related
pollution.

We've got access to appropriate facilities, materials, and equipment---
not of the highest caliber, though I'm working on that, but certainly
workable. It's one of the relative advantages of working under the
material and social conditions of a university. On the other hand we
could never pull off what AQE participants have here! So hopefully
there's a symbiosis.

I'm volunteering to put in the time and buy materials for testing,
analysis, and model fitting/diagnostics of the assembled modules. That
may be better than testing the sensors and modules separately, though
we can always swap sensors between modules to get data on that, or
test sensors with minimal support circuitry.

Before I saw your message here, I posted about it to this thread:

http://groups.google.com/group/airqualityegg/browse_thread/thread/56434c9e0af8b3bb

From our perspective in public health, even if we COULD measure NOx
with infinite precision, it's probably not precisely what we need to
care (most) about. Pure NOx doesn't have extremely toxic effects at
ambient levels. But, it's a decent marker for the stew of aerosols
that does cause health effects, and as such, it's very useful for
developing inference about health effects and about exposures to
traffic-related pollution in general.

Cheers, David

[David Holstius]

One more brief comment, IMHO Neil you are on the right track with all
of this and I'm glad you're here and pushing for good engineering
process as a way to increase the relevance and credibility of the Egg
readings. Everyone here seems to want value out of the process, people
are just emphasizing different kinds and different ways of getting at
that. It's really great to see.

[NeilH]

Hi David

Thanks for the comments - NO2 is just a suggestion for people who might have experience to be able to describe it - and because I've been digging into the NO2 data sheet.

From the point of view of an equipment maker - the biggest risk is how much people want the equipment and of course the whole maker community is facility more people get involving - which fundamentally brings down the total cost of instrumentation for everyone.

My perspective as an engineer who has done it for other sensors, that the process of getting raw data from a sensor and translating it to measurable units is an engineering challenge - the rubber meets the road - how to make meaningful use of the data generated.

You have to have been through the process to realize how tortuous it is - and since its manufacturing, it needs to be repeatable at a low cost point for the final costs to be kept low as well.

 

The typical steps that happen with new equipment are 

Concept - Air Quality measurements (it could be just another Arudino based experimenter kit with add what ever sensor you want)

decode systems functions (In this case Arduino based shield with internet reporting & viewing)

make individual subsystems - hardware, firmware and pachube interfaces

verify system function - from sensing (typically in reference units) to presentation (in this case on the wwww wonderful world wide web)

validate instrument - put in final context with all associated natural phenoma (weather, temperature, ants,  people kicking  ... whatever :)

So, it sounds like  you are suggesting you have the equipment to characterize the AQE output sensor data - whatever that output may be. 

Which would be fantastic and it would be fascinating to see.

There is a document where some analysis was done just on NO2 and it might be a reference. "NO2 Sensors Report"  http://airqualityegg.wikispaces.com/Hardware-Sensors

What might also be great, is if you can document simple safe ways of verifying "calibrated sensors" - develop a  standard operating procedure (SOP) for anyone to do.

For instance - can the sensors be placed in a tedlar bag and a source of "Synthetic Air" introduced - which is the manufacturers recommended first step in the calibration - determine the sensors base resistance in "Synthetic Air". 

According to the MiCS-2710/2714 data sheet base resistance (sensing resistance in the reference synthetic air) R0 can vary from 0.8Kohms to 8Kohms. 

Then in say another tedlar bag with a connecting pipe with a valve - add a known quantity of gas under test, to bring it to an Air Quality threshold of interest, for NO2 this could be100ppb - and another measurement taken. 

The MiCS-2710/2714 says the resistance is proportional to the presence of NO2 and the slope of that variation is non linear and can be between 6 and 100.

Maybe its pie in the sky idea to look for a field based SOP as well. Just an idea.

[Cesar Garcia]

Jorge, who is also on this list has offered to calibrate one of the assembled prototypes also, so let's hope we can share details and data.

BTW, has anyone asked Libelium about the calibration process for the 2710 on their boards?

Best,
Cesar

[David Holstius]

A cheap and easy SOP would field calibration would be extremely
valuable. I'm not sure how to do it for NO2, but for PM there's a clue
offered by the zero check procedure for the UCB Particle Monitor,
intended for use in developing regions where PM exposures are very
high due to widespread use of biofuel cookstoves:

http://www.scribd.com/doc/48104650/UCB-Particlemonitor-SOP

Page 5: the zero check is just to leave the particle monitor in a
Ziplock bag, undisturbed, for 40 minutes. If you had a ready source of
filtered air, this could make for a faster zero check, so maybe
something like a HEPA filter from an air purifier or filtered vacuum?

Once the offset is established, and assuming it's stationary, then
obviously (as you know) the question is the gain. The potential
density of a deployed network based on devices like AQE suggests a
different problem than one-at-a-time factory calibration. Balzano (of
CENS) and Nowak call this "blind calibration" of sensor networks. Some
theoretical and simulated results are here:

http://sunbeam.ece.wisc.edu/publications/bcbook.pdf

That general discussion is predicated on the process of interest being
outdoor/ambient air, though I don't see why it couldn't be applied to
microenvironments indoors as well.

[Victor Aprea]

Something nobody has talked about so far is what the operator interface for calibration might be like. I think it would have to involve the use of computer with a serial terminal and a command/response type of protocol that lets the operator identify a sensor that is on the bus to calibrate and then inform that sensor what the current gas concentration is. Development of a standard operating procedure for field calibration would be an excellent way to tease the details of that operator interface and what that might mean to the sensor modules and Nanode software design.

Regards,

Vic

[NeilH]

On Sunday, June 3, 2012 8:54:50 AM UTC-7, David Holstius wrote:

....

Once the offset is established, and assuming it's stationary, then
obviously (as you know) the question is the gain. The potential
density of a deployed network based on devices like AQE suggests a
different problem than one-at-a-time factory calibration. Balzano (of
CENS) and Nowak call this "blind calibration" of sensor networks. Some
theoretical and simulated results are here:

  http://sunbeam.ece.wisc.edu/publications/bcbook.pdf

Hello David

Fascinating. Blind Calibration.

I read it as possibly solving problems as associated with drift in sensors in the field, they differentiate between uncalibrated sensors (with manufacturing offsets applied) and raw sensors measurements

They identify conditions which if raw sensor output meet could apply more generally

Some excerpts  with an attempt at translating it to a gas sensor..

blind calibration -  a novel automatic sensor calibration procedure that requires solving a linear system of constraints involving routine sensor measurements. By “routine” we mean that actual signal measured by the sensor network is uncontrolled and unknown. 

"sensors slightly oversample the signals of interest," - seems to imply that a physical array of sensors all measuring "over sample" the same gas "signal of interest"

"nor a dense deployment is required" - they suggest a matrix of 8 x 8 sensors, but later use 9 temperature  sensors which have been linearized,  though the general case is "from n sensors lie in a subspace of n-dimensional Euclidean space"

so my maths gets a bit rusty.

"We assume a linear model for the sensor calibration functions. This means that the sensor readings are calibrated up to an unknown gain and

offset (bias) for each sensor, possibly after applying a suitable and fixed transformation to the raw sensor readings, e.g., taking the logarithm or

applying the original factory calibration transformation."  - most sensors are not linear across their range of interest, but maybe there are linear relationships that can be determined from knowledge of the materials (!!) - however drift can be linear, as it applies to all parts of the sensor (?).

"The Nyquist theorem dictates a minimum spacing between sensors in order to adequately sample a bandlimited signal.

 If sensors are spaced more closely than the minimum requirement, then we are “oversampling” the signal."

That seems to suggest if the spacing (mechanical design) between the sensors is sufficient to so the measurement error (gasseous distribution and electronics measurement noise) is less that a pre-determined minimum - then the ppb in the gas signal can be determined to be oversampled

(mental note - some needs to understand how close the gas sensors need to be and what level of measurement signal is needed)

Moving into more detail

We can summarize this for all n sensors using the vector notation

x = Y a + b (1)

where Y = diag (y)   - the oversample  measurements from 'n' sensors.

The blind calibration problem entails the recovery of 'a' and 'b' from routine uncalibrated sensor readings such as y.

now fast forward to sect 6.2.1 Calibration Dataset

"the conclusion was drawn that after the factory-supplied calibration was applied to the raw sensor measurements,

the sensors differed from the reference thermocouple linearly i.e. by only a gain and offset."

and they use 9 temperature sensors for the calibration dataset.

After that they apply it to 26 distributed temperature sensors

and then I'm afraid I can't follow how they are using the data. Plus I'm running out of time.

Sect 8 Extensions and Future Work

"There are many issues in blind calibration that could be explored further.

The two main areas ripe for study are the choice of the subspace P and the implementation of blind calibration. 

 ..

How to choose the subspace when faced with a sensor deployment where the true signals are unknown is an extremely important question for blind

calibration.

and

9 Conclusions

The problem of sensor calibration is central to the practical use of sensor networks. 

..

We have demonstrated a working implementation on simulated and real data, which uncovered interesting relationships between implementation and blind calibration performance. 

Overall, we have demonstrated that blind calibration has great potential to be possible in practice, and we feel that the proposed formulation merits further investigation.

Gotta leave it at that .. but be interested as to what other people make of it, and I'm wondering why they don't use the raw sensor data from a temperature probe - and temperature after all is the most measured physical parameters and one of the most easily to relate to, and one of the most desirable to get a good accuracy on. 

[NeilH]

Blind Calibration is intriguing and I've been thinking about the paper. I've started a new thread on it "Blind Calibration and measurement theory" for further discussion.

[Michael Heimbinder]

We setup a low-cost lab for evaluating low-cost gas sensors. I thought some egg owners out there might be interested in our methodology and findings, see http://www.takingspace.org/evaluating-low-cost-gas-sensors/

-Michael-

[NeilH]

On Thursday, October 17, 2013 10:38:16 AM UTC-7, Michael Heimbinder wrote:
> We setup a low-cost lab for evaluating low-cost gas sensors. I thought some egg owners out there might be interested in our methodology and findings, see http://www.takingspace.org/evaluating-low-cost-gas-sensors/
>
> -Michael-

Wow .. great to see your article and range of manufacturers sensors you tested.
Your comments on the Mic2710 "high out of the box variability" fits with what the data sheet frames.
It is one of the challenges for the design to be able to have the hardware non volatile EEPROM/Flash capability and software algorithms to incorporate each unique Mic2710 + calibration procedures into delivering measurements that are calibrated units.
Thanks for sharing it.