Lukashova Lidiya Sergeevna
Possibility of Locating Heat Source of Spontaneous Combustion of Coal by the Electric Potential
The Individual Task
"The computerised subsystem of diagnostics and monitoring
of a condition of the mountain machine
(on an example of coal-mining combine 1KDK500)"
DISCRIMINATORY MINE FIRE SOURCE DETECTION
Product-of-combustion sensors were used to discriminate mine fire sources
of coal, diesel fuel, electrical cable insulation, conveyor belt, and
nuisance emissions from acetylene torch cutting operations in experiments
conducted in the National Institute for Occupational Safety and Health
(NIOSH), Pittsburgh Research Laboratory (PRL) Safety Research Coal Mine
(SRCM). The sensors consisted of CO, ionization and optical smoke, and
metal oxide semiconductor (MOS) sensors. Metal oxide semiconductor and
smoke sensors demonstrated an earlier fire detection capability than a
CO sensor. This capability was of particular significance for a smoldering
conveyor-belt fire in which the optical visibility was reduced to 1.52
m with an increase in CO of less than 2 ppm at a distance of 148 m from
the fire. An application of a neural-network program to the sensor responses
from each type of fire source resulted in correct classifications of coal,
diesel-fuel, cable insulation, belt, and metal-cutting combustion with
a mean of 96% of the in-mine test data correctly classified. In a battery
charging building, a fire sensor configuration consisting of an ionization
type smoke sensor and a MOS, NOx -sensitive sensor was demonstrated to
be capable of discriminating a coal fire from diesel equipment when H2
from a battery charging operation saturated the CO chemical sensor cell.
NEURAL NETWORK ANALYSIS
A
neural network analysis was applied to the classification of fire sensor
responses to differentiate between possible fire sources. In this neural
network program, the time dependent fire sensor data were compared to
the nonlinear approximations generated by the neural network until adequate
approximations for correct classifications were obtained through corrective
iterations. The input layer of neurons contained the experimental sensor
data. The output layer of neurons contained the fire source classifications
generated by the neural network. Between the input and output layers were
two hidden layers of neurons or process elements (PEs). The inputs to
the hidden layers of neurons were multiplied by weights, summed, and processed
through a bounded, nonlinear activation function. In the training phase
of the neural network, the output classifications were subtracted from
the correct classifications and the differences, or errors, were used
by a back propagation method, which was a modification of the gradient-descent
search technique, to adjust the values of the weights until a sum of the
errors was adequately reduced over a reasonable time interval. For the
sensor data analysis considered here, the neural network software package
entitled NeuroSolutions for Excel from NeuroDimension, Inc. was used.
Neural Analysis
The neural network application was restricted to the in-mine experiments
listed in Table 2 and an acetylene cutting experiment. The complete set
of sensors for which measurements were available for the fire experiments
listed in Table 2 were CO, FA, FB, SA, and SB. In order to use the neural
network program, the data for each experiment were prepared in files with
the fire sensor signals normalized to their ambient background signals.
The responses of FA and FB were nearly identical. The training of the
neural network was accomplished with the five sets of sensor data from
coal, diesel-fuel, electrical-cable insulation, and conveyor-belt fires,
which are the fires of experiment nos. 1 to 4 in Table 2, and an acetylene-torch,
metal-cutting experiment. Seven data inputs were processed from the sensor
data to classify the five fire types. The inputs, which include time and
multiplicative combinations of the data from four of the sensors but excluding
sensor SB, were determined by trial-and-error to be the most suitable
inputs for accurate classifications. The size of the training data sets
ranged from 85 to 991 exemplars, or time samples, of the four sensor inputs
and two functions of the sensor inputs with the total size of the training
set being 2,988 exemplars. The two functions of the sensor inputs were
the product of CO and the average of FA and FB, and the product of CO
and SA. This was determined by trial and error. Time zero at the beginning
of each data set corresponded to the first sensor alarm for each type
of fire. Sampling by the sensors occurred at two-second intervals.
Various neural network programs provided in the package by the vendor
were applied to the data in attempts to successfully classify the fire
types. A two-hidden-layered perceptron network with momentum-back propagation-of-error
algorithm produced reproducible results. The first hidden layer consisted
of twenty neurons, or process elements (PE), and the second hidden layer
consisted of ten PEs. It was discovered that the testing results were
reproducible even though the initial weights between the PEs were assigned
randomly before each training calculation. The activation function used
in the hidden layers was the hyperbolic tangent function with the output
layer using a softmax classification function. One thousand epochs, or
iterations, through the samples were performed with error correction after
every epoch. The minimum squared error achieved after one thousand epochs
was 0.0012.
For testing the neural network, seven data files were presented to the
trained network. These files included experiment nos. 5 to10 in Table
2 and one acetylene torch metal-cutting experiment. The number of testing
exemplars in each file ranged from 121 to 1,854 with the total size of
the testing set being 4,255 exemplars. Two coal and two diesel-fuel fires
were included in the set of testing files.
The average
correct classification of the fire sensor data for the seven tests in
table 3 is 96%. The minimum value of 86% for a single experiment is not
unreasonable. Evaluation of experiments 11 and 12 could not be made with
the neural network program because experiments using materials similar
to those of BELT2 and BELT3 were not available to include in the training
set.
DonNTU | Master's Portal | Searching System of DonNTU