Real-time neural network application to mine fire – nuisance emissions discrimination

: The National Institute for Occupational Safety and Health (NIOSH) implemented a real-time neural network system which can discriminate mine fires from nuisance diesel emissions as part of an atmospheric mine monitoring system in NIOSH’s Safety Research Coal Mine. The real-time response of a neural network to fire sensor outputs was demonstrated for coal and belt combustion in the presence of diesel emissions. The fire sensors consisted of an optical path smoke sensor, a carbon monoxide (CO) sensor, and two types of metal oxide semiconductor (MOS) sensors. The real time neural network was trained with coal, wood, and belt fire experiments with and without diesel emissions background. The trained neural network successfully predicted mine fires with these combustibles in the smoldering stage prior to the onset of flames. welding, and its susceptibility to interfering gases, such as H2 from battery charging operations. Smoke sensors, which are either optical or ionization, are responsive, stable, and easy to check. The ionization smoke sensors are usually based upon the decay of an alpha emitter such as Am or a beta emitter such as Kr. A disadvantage of an ionization smoke sensor is the requirement for a license for the radioactive source dependent upon the strength of the source. Optical smoke sensors can be either a point or a path type. The point type sensor is based upon optical scattering, and the path type is based upon optical obscuration. An advantage of the path sensor is its ability to probe the cross section of an entry. A disadvantage for a path optical smoke sensor is the optical path distance which can create a problem for its installation in a mine. Nuisance emissions which can affect smoke sensors include dust, water vapor, and particulate emissions from diesel equipment and flame cutting, and for the case of an optical path sensor the potential blockage by personnel and equipment. Optical smoke sensors, which are generally in the infrared optical range, are very responsive to larger smoke particulates produced by smoldering combustion, and ionization smoke sensors are generally more responsive to smaller smoke particulates produced by flaming combustion. The demarcation size is about 0.3 micrometer. Because of the optical sensor’s response to larger particulates, it will be less responsive to smaller diesel particulate emissions. Another class of sensors which respond to mine fire gases are MOS sensors. These sensors are very responsive, but not very selective in their response. Their responsiveness can be ordered in terms of their resistance change to various oxidizing and reducing gases. Methods to vary their selectivity include the addition of a catalyst, sensor element grain size selection, heating of the sensor, and filters. Another advantage of the MOS sensor is its compact size. Its cylindrical structure has a height and diameter of about 1.3 cm. A disadvantage of a MOS sensor is the high temperature, 250 to 380 o C, of the sensing element. A MOS sensor which is sensitive to NOx has an increase in its electrical resistance associated with adsorption of oxygen due to electron transfer from the surface element to the oxygen and the buildup of a positive space charge on the element surface. The presence of deoxidizing POC gases will remove the adsorbed oxygen and result in reduced resistance. To be useful, the MOS’s sensitivity to NOx must be sufficient to overcome its sensitivity to CO in diesel emissions, otherwise the discriminating capability of the sensor will be compromised. Generally there are gas concentration ranges over which the sensor’s resistance element will respond with a power dependence upon the adsorbed target gas concentration. For an extended range of concentrations this nonlinear logarithmic dependence makes the prediction of the net effect problematic, and direct experimental verification is required. In a refueling area where rapid flaming combustion most likely will occur, optical spectral emission sensors could be useful. In this case the distance between the source fire and sensor will be relatively short, and infrared signal attenuation by smoke will initially be minimal. The ordinary use of fire sensors depends upon defined alarm values. Fire sensors for in-mine use have established alarm values (Code of Federal Regulations, 2001). These include a CO alert value of 5 ppm above ambient, a CO alarm value of 10 ppm above ambient, a smoke optical density of 0.022 m, and a temperature of 165 ° F. 3 MINE FIRE DETECTION METHODS There are three approaches for utilization of sensors for mine fire detection in the presence of nuisance emissions. One approach is to increase the sensor alarm and alert values to compensate for nuisance emissions. Without an exact knowledge of the nuisance emissions’ gaseous and particulate concentrations, which will be variable in response to the operation of the emissions producing equipment under variable power loads, early and reliable fire detection will be compromised. For example, this approach would be expected to miss early smoldering stages of a mine fire. A second approach is rule making. This approach would rely upon an understanding of the generation of emissions from nuisance producing sources and combustion sources. For example, this approach could examine rates of change in a measurable emission, such as CO, so as to differentiate CO emissions from a diesel and from a mine combustible. However, the rate of change of a fluctuating time variable quantity will fluctuate considerably and can produce indistinguishable events. Another example is to rely upon the measured historical relationship between two emission components, such as NO and CO from a diesel engine. This approach can depend upon the growth rate of the fire (Edwards et al., 1999). A third approach is information processing. This approach is useful when multiple type sensors with nonlinear responses to various POC are used. In this approach a relationship between the classification of an event and the sensors’ inputs can be established with a method which recognizes patterns for known data sets. An advantage is that a number of mine firenuisance emission experiments can be conducted to establish the functional relationships. One such information processing approach is a NN approach. Based upon these relationships classifications of unknown events similar to those for which the relationships are developed can be made. These classifications can be quantified with a probabilistic interpretation.