Gas Detection and Alarm Systems: Why a Good Calibration Fit Needs Independent Validation

Temperature and humidity compensation can improve a methane sensor's interpretation, but a close fit to calibration data is only part of the evidence. The useful question is whether the relationship still works when the sensor encounters a separate measurement sequence. That question becomes important when software claims are used to support gas-monitoring equipment selection.

A flexible model can reproduce the particular correlations present in its fitting data. If temperature, humidity and methane concentration change together in that experiment, it may learn a relationship that is less reliable when they vary differently. A high coefficient of determination on the fitting set therefore does not establish dependable behaviour across the complete operating range.

An independent evaluation should define the environmental and concentration conditions being challenged. It should also identify which parameters were fixed before the evaluation, rather than adjusted after looking at the new results. This allows a buyer to distinguish a genuine validation from another round of calibration.

Mitchell, Cox and Lewis studied a low-cost methane sensor for low-concentration environmental monitoring and compared regression models that included temperature and humidity. Their separate decay experiments exposed limitations that were not apparent from calibration fit alone, including overfitting in one model. The reproduced multi-panel figure shows estimates from the different models during those decay experiments. [1]

Methane estimates from four model equations applied to two independent decay experiments. Original equation identifiers, axes and units are retained to preserve the comparison. Source: Hazel Louise Mitchell, Simon J. Cox and Hugh G. Lewis (2024), Calibration of a Low-Cost Methane Sensor Using Machine Learning, Figure 12. CC BY 4.0. Selected figure/panel extracted by rectangular crop and resized; no data, scales, units or necessary legend changed. Source figure number remains in the external caption.

The work investigated low methane concentrations and a particular sensor module. It is not a certification study of a complete combustible-gas alarm. Its model equations, estimated concentrations and errors must not be presented as Apexmeter detector specifications or used directly as alarm settings. The warm-up plot on the cover also belongs to that sensor and experiment.

Warm-up, supply conditions, contamination, sensor ageing and the gas matrix can affect response. Temperature and humidity measurements themselves have location and response characteristics. An algorithm trained with one arrangement may not represent a different enclosure or sensor replacement without further evidence.

For a Southeast Asian procurement, ask which humidity and temperature combinations were actually tested, including recovery after disconnection or power cycling where relevant. Determine how the device reports a condition outside the model's validated range. A numerical output should not silently imply a validated measurement when the supporting assumptions no longer apply.

A useful sensor study can improve the buyer's questions without establishing suitability for a protective duty. The complete detector and alarm chain need evidence appropriate to their intended role and the applicable local requirements. An environmental methane research result should not be translated into a safety claim merely because both applications involve methane.

Request validation reports tied to the selected sensor, enclosure, electronics and firmware. Define how replacements or software changes are reviewed and how the supported maintenance procedure restores confidence in measurement. This gives the buyer a clear way to assess compensation claims while retaining independent evidence as part of the selection decision.

Send the following project inputs when requesting a configuration and evidence review:

Cover image: Warm-up behaviour of the study methane-sensor module; the paper observed stabilisation after approximately four hours under its experimental conditions. Hazel Louise Mitchell, Simon J. Cox and Hugh G. Lewis (2024), Calibration of a Low-Cost Methane Sensor Using Machine Learning, Figure 6. CC BY 4.0. Selected figure/panel extracted by rectangular crop and resized; no data, scales, units or necessary legend changed. Source figure number remains in the external caption.