NIOSH Mining Safety and Health Research

To protect miners from lethal smoke and combustion gases associated with fires, research is being done to improve fire detection technology and strategies for the early and reliable detection of underground mine fires. The increased use of diesels in underground coal mines increase the probability of mine fire sensor false alarms. Mine fire sensors are responsive to the same carbon monoxide (CO) and smoke particulate emissions from both diesel equipment and open combustion. Other underground mine nuisance alarm sources include hydrogen (H2) emissions at battery-changing stations, which can interfere with CO sensor chemical cells, CO and smoke particulates from welding and flame cutting activities, and airborne particulates associated with rock-dusting operations. Repeated false alarms can increase the likelihood that miners will ignore all alarms. The presence of various nuisance sources with their characteristic emissions suggests the need for a combination of multiple sensor types to differentiate the various nuisance events from a mine fire.

Mine fire detection experiments were held in the Safety Research Coal Mine (SRCM). They included isolated coal, conveyor belt, and wood fires, as well as these fire sources in the presence of background diesel emissions. The strategy was to determine the optimum selection of multiple sensor types that could provide early and reliable mine fire detection. Diesel and flame cutting emissions include nitric oxides (NOx). Their signatures were demonstrated to be distinguished from fire products of combustion (POC) by a metal oxide semiconductor (MOS) sensor that has a bimodal response to NOx and hydrocarbon POC. The fire sensor strategy must also be able to differentiate the cross-sensitivity of CO sensors to H2 produced by battery-charging operations. This defines a need for a smoke sensor. Since a CO measurement is necessary quantify a fire´s intensity, a CO sensor remains a primary sensor. The research evaluated the response of as many sensors as feasible for each mine fire experiment. These sensors included CO, MOS, and ionization and optical smoke sensors. To establish rules for mine fire detection from the responses of the sensors to a variety of fire conditions and nuisance emissions, a neural network approach that classifies events based on a training set of known events was used. For the training set, sensor experimental data from a coal, a conveyor belt, and a wood fire in the presence, and in the absence, of diesel emissions were used. The optimum sensors determined from a trial-and-error approach were: CO, optical smoke, an MOS sensor responsive to POC, and an MOS sensor with a bimodal response to POC and NOx. An outcome of the neural network application is the classification of the real-time event into clear air, diesel emissions, and mine fire. A probability is associated with each classification as a function of time. Application of the trained neural network to coal, conveyor belt, and wood fires in the presence of diesel emissions successfully predicted the occurrence of a fire in its smoldering stage before the onset of flaming combustion. Based on the success with postexperimental analysis, a real-time interactive computer program was developed to acquire the selected sensor data and prepare it in a normalized format for real-time application of the trained neural network. The successful detection of combustion in the presence of nuisance emissions was demonstrated with a real-time application to coal combustion in the presence of diesel emissions in the SRCM. The neural network application detected the fire in the smoldering stage 13 minutes before it transitioned to flaming combustion. Because of the fluctuations in the diesel CO emissions, the CO sensor did not indicate a fire event until flaming combustion.

The advantage of a multiple-type fire sensor methodology is that sensors can be replaced with more advanced sensor technologies as they become available. The presence of additional combustible materials in underground mines will require an extension of the methodology with an expanded database for a neural network training set.