"Network Calibration" project kick-off
Martin Dittus
after a few discussions with Ed we realised he's already very busy with overall AQE project coordination, so I offered to coordinate the "network calibration" aspect: using the size and distribution of a large sensor network to improve on the sensing quality provided by its constituent devices.
Over the next few months we want to form a good understanding of our network calibration limitations and opportunities; to review a few approaches, and then design a first concept. It's exciting to ponder what we can do with a collective sensor network of dozens or hundreds of participants; and to then consider the prospect of thousands and millions of nodes :D
Picking up on a few discussion threads I suggest four phases/strands:
1. Collect known system limitations (how good/bad are our sensors really?)
2. Collect recommendations for metadata to collect ("traceability")
3. Review potential network calibration models, pick one
We appear to have quite a few experts in the group already, particularly when it comes to sensor quality, common calibration practices, metadata questions, data quality requirements for further usage, etc. Your suggestions and critiques so far have been excellent, and your offers for help are much appreciated. It would also be good to find sensor network experts with advice on how to model network coverage and measurement uncertainties.
(Based on a conversation I had with Neil Hancock it seems there's also a discussion to be had on whether "network calibration" is the right term to use for this; but in the interest of moving things forward I'd be happy to stick with it for now and change it later if needed.)
Have I overlooked anything? Does this seem worthwhile to everyone?
Briefly about myself: I'm a postgraduate student at UCL researching collective sensor networks, and intimately familiar with questions of data cleaning, metadata quality, social data, and data modelling problems and techniques. I'm very impressed with this young community, and with the quality of our discussions!
Looking forward to tackle some hard problems with you guys,
Martin Dittus
Martin Dittus
- Network density: encourage overlapping sensor placement
- Randomised controls: calibrate a few devices before sending them out, keep track of them
- Redundant sensors: ship some devices with 3-5 uncalibrated sensors instead of just one
- Augmentation with qualitative ("social") data
- Correlation with data from other sensor networks
- Calibration at home (e.g. as local community project)
- Careful choice of sensor components
Hope this gives you some inspiration for things to ponder… I'm skipping lots of details here of course, but you can see that there's quite an interesting spread of things to investigate.
m.
NeilH
Hi All
I’ve a basic question – if I purchase an AQE will it be able to measure specific gas concentrations that can be used in an Air Quality matrix as defined by Air Quality Standards.
I need to ask this basic question because the AQE description lack meaningful specifications. The AQE description – in my opinion - is using flowery language that is not helpful in defining exactly what the sensors are designed to do.
From my experience data processing is in layers – the fundamental physical conversion layer needs to be validated first and verified. If its not specified it almost certainly is likely to be flawed. If the physical data collection layer is flawed the integrity of upper processing layers are likely to be highly suspect.
Ed Bordens statements that calibration is the wrong question suggest he just doesn’t understand real world sensors. I’m happy to be educated if I have missed something. But repeated referencing the same statement is not persuasive.
I’m an electronic/software design engineer who has designed other transducers to use physical sensors to the limit of their specification.
Designing a circuit to transform a physical quantity (NO2 CO O3) needs to follow some simple rules.
There needs to be a physical sensor that transforms the concentration of gas to a measurable electrical value.
This is then converted through an electronic circuit to the digital world (Analog to Digital Conversion)
An algorithm running on the processor then typically “digitally signal processes” into a useable trace-able values.
I’m not an Air Quality expert, so I’m using thresholds from a Sonoma Technologies document called “Desired Characteristics and Information about Major Air Pollutants for Device Manufacturers to Use in Creating Instruments for Non-regulatory Monitoring (e.g., Citizens, Schools, NGOs).”
If someone has other specifications please share them.
The following thresholds fall out of the table
CO to 9ppm with an accuracy of 1ppm. Normal atmospheric ranges are up to 20ppm
NO2 to 100ppb with an accuracy of 1ppb. Normal atmospheric ranges are up to 400ppb.
As of a couple of days ago the NO2 sensor was specified as the MiCS-2710. So as I read this data sheet the sensor is not capable of monitoring to the above specification. It has a Sensor response specified as from 0.2ppm to 1.5ppm (200ppb to 1500ppb)
The MiCS-2714 does seem to have the physical range of NO2 concentrations. Its sensitivity is 10ppb to 900ppb
Since this is an Open Hardware Design – I’m offering my input to get a sensor that is understood – if I’m missing something – please lets have a system design document so its is clear.
The CO sensor MiCS-5525 does seem like it can meet the above specification. The sensor characteristics show the sensor capable 1-400ppm
However it’s a non-linear relationship, and the lower sensitivity of 1 to 10ppm is on the flattest part of the curve – which could be the most challenging to measure.
All the sensor specifies that the active measurement be compared against a synthetic air measurement (presumably down at the factory).
All the sensors define the range they measure could be different for each sensor – so is it measuring 1 unit or 100units – has to be determined for each individual sensor.
Look forward to hearing sensible responses to these basic questions
Cesar Garcia
I've been checking e2v MICS-2710 datasheet (http://www.e2v.com/e2v/assets/File/sensors_datasheets/metal_oxide/MICS-2710.pdf) and the NO2 detection range appears from 0.05 to 5 ppm.
--
Cesar García - @elsatch
Ando con encolamiento para responder correos y los proceso lunes, miércoles y viernes. Si es algo urgente/rápido contáctame por Twitter. Gracias!
NeilH
Hmm Yes interesting, thankyou for looking at the data sheet, – I was only looking at the MiCS-2710/4 graphs. (I’m on vacation at the moment, and don’t have ability to print and compare data sheets carefully)
So looking at the differences between MiCS2710 and 2714 they are different in package types, but otherwise appear incredibly similar.
What is intriguing is that the graph of Rs/Ro cover different ranges. However since the 2714 is so similar it suggests that at the more sensitive end the range – that of interest in Air Quality thresholds - is non-linear. The Application Note datasheet MiCS-AN2 FAQ states
“As the characteristic curves of the sensors are non-linear, e2v recommend the use of multi-point calibrations, covering the whole range of the application in terms of temperature, humidity and gas concentration. Once the set of calibration parameters is obtained, the data can be interpreted using polynomials or curve fitting techniques. “
Also it seems to me the “NO2 Detection Range” of FS min=0.05ppm to max=5ppm needs some careful analysis. I hate to do too much of a casual analysis on it – but it seems to me that for some individual sensors may only measure up to 50ppb and then max out – flat line – and then other sensors could measure up to 5000ppb before flat lining.
So the meaningfulness of the measurements “accuracy” comes at a cost – and the question is can a sweet spot be found that is MINIMAL calibration for a defined range that is sufficiently accurate for Air Quality thresholds.
The data sheet doesn't state any long term drift that I can see of.
Its very painful that the designers are not stating what is the range they are aiming at.
Martin what sort of information do you think you need to be pull any “trend” information out of the production sensors.
What assumption do you make about geographically linked sensors - and how close do sensors have to be to be able to make any assumptions that they are measuring something similar. Or another way of asking how well does air mix up pollutants and distribute them?
Nafis
Victor Aprea
Martin Dittus
m.
Cesar Garcia
Last week I sent an email regarding an air pollutant city simulator, and how the eggs could be used to validate model in quasi-real time. On 22nd May, we'll have a big debate on Madrid air quality at Medialab-Prado and we've been very lucky to get the air quality department director, Roberto San Juan, to attend and present their approach. They have been involved in environment simulations for the city council for ten years at least. We will talk about air quality egg project and some of the possible usages. I think it could be a very good opportunity to ask some interesting questions, so the thing is: Is there any specific topic or (hard) question would you like us to ask him? I'll do my best the get the answers and bring them back here.
Looking forward to hear you proposals,
César
P.D Check the website for their current work (http://artico.lma.fi.upm.es/) and to try to focus questions. I assume he won't be able to answer specific model sensor questions (MICS-2710 vs MICS-2714, for example).
Thanks!
m.
NeilH
Cesar Garcia
Martin Dittus
I'd be very interested in hearing some details of how they handle data quality, particularly with data coming from a range of sources. What kinds of annotations do they produce per data source to describe uncertainties; the various ways in which they assess and quantify uncertainties of individual sensors and sources (e.g. measurement errors, device limitations, fluctuations caused by environmental factors, etc.) But also how they then model systemic data quality and uncertainties: do they attempt to quantify the accuracy of their models? Do they have any correction methods in place to adjust for e.g. temperature fluctuations? And if so: could we get access to their documentation?
For our "prediction model" group it would also be interesting to ask him about references for monitoring and forecasting methods. Essentially: once they have prepared the data, what kinds of data models do they build on top of it?
Let me know of you'd like me to elaborate any of the above. Happy to discuss off-list as well.
This could be a very interesting contact, pity I can't join you :)
m.
Martin Dittus
Let me know if all of this is too much, I can summarise.
m.
NeilH
Martin Dittus
On 18 May 2012, at 19:25, NeilH wrote:
> To quote from the paper - pg319 - and this I believe applies more so to a gaseous regime
> "several factors make the calibration challenging.
> Firstly, the calibration can only occur when the ground truth node and the uncalibrated node are experiencing identical sensing environment.
> This is necessary because the comparison between calibrated and uncalibrated data is only meaningful when the same input to the sensors is applied.
> Secondly, the calibration rendezvous is complicated by the existence of the sensing factor.
> The sensing factor is identified by the tendency of a physical phenomenon to be localized to a small region around the entity taking the measurement"
And should we find out that calibration-by-neighbours is impracticable for most cases then there's a whole range of other options, as I suggested above.
(If I may point out, you have a habit of hijacking threads :) Could I ask you to start new threads for new topics? This makes it much easier for everyone to follow the various debates, and to browse our email archive. In many email clients this is as simple as changing the subject line.)
m.