AbstractVery effective detection of spontaneous combustion or open fires in coal mines can be provided from calculations based on measurements of a few parts per million of carbon monoxide (CO) in the ventilating air, even in the presence of CO from shotfiring, diesels and other sources. A project to develop automatic alarms is described.Forecasting/smoothing methods were abandoned in favour of the development of a model of normal conditions as a Markov process, with an associated sequential hypothesis test. Various approximations and assumptions allowed the derivation of algorithms to suit different circumstances, including both continuous and intermittent sampling of CO. In each case, the calculation of alarm levels for individual locations was an important part of the algorithm.The model allowed the most to be made of a minimal knowledge of rare events (i.e. heatings). The‘quick and dirty’algorithms proposed are simple enough to allow an 8 bit micro to monitor 20 points while simultaneously revising alarm levels; yet being derived from a good model, they are effective, well behaved and adaptable.