ASIC discussions look at AI vs non-AI calibration

So how do you do it? Many presentations at the recent ASIC conference revolved around calibration of small sensor air quality systems, including that given by AQMesh Technical Business Development Manager, John Downie. Offered the opportunity to poll the ASIC audience, we chose to ask delegates ‘Please indicate which of these calibration methods for small sensors you would consider for your small sensor network’. Whilst numbers responding were not huge, ‘periodic co-location of each instrument with FRM/FEM creating seasonal correction factors’ came out top, indicating a thorough – if labour-intensive – approach is generally most common and taken as the ‘norm’.

So what if you don’t have the ‘boots on the ground’ to carry out the many instrument movements that would be necessary with a network of any size? Is there a short-cut? Can AI help? The poll did show that about a quarter of respondents were considering ‘desk-based calibration, using machine learning or AI’, about twice the proportion thinking about ‘desk-based calibration using network calibration method without machine learning or AI’.

We tried using AI on electrochemical sensor output a few years ago but found that – at the time – the machine learning was very good at latching onto interfering factors, such as temperature, and less good at finding the weak (valid) signal in amongst the noise. We were greatly impressed by the later (non-AI) network calibration work done by University of Cambridge, our partners in the Breathe London pilot, and have developed our own approach. This series of repeatable calculations can be applied to a network of five instruments upwards and we feel there are some clear differences (advantages, of course) comparing the AQMesh approach to local calibration with AI-driven approaches:

The fundamental issue with machine learning and AI being used either for compensation of sensor interferences or for calibration adjustments is that their whole premise is to “evolve”: they spot differences and change to account for those changes. This means that over time the results may be closer to reference, but they will never follow the same process for adjustment as they did previously. This raises significant traceability questions about when processing changes and makes it impossible for these methods to become certified to a standard which require a fixed processing of inputs, managed by a version number.

It’s worth noting that PAS 4023 (Annex D) distinguishes between sensor systems that are ‘self-contained’ – with readings derived by the system alone, using a series of repeatable calculations – and those that require training against a reference station. The PAS also emphasises how important it is that individual sensor systems should perform very similarly to one another, so that remote comparisons across a network can be made with confidence. Another reason why a proven, global correction algorithm used by every instrument has the edge over individual site-based AI. Strong inter-instrument comparability also means that a network can be meaningfully compared against itself, in the absence of a reference station, as offered by AQMesh’s network normalisation option.