station. After delivery, the coal is often piled up in stock at random and blended
from all kinds of sources (Chen et al., 2010; Qiu et al., 2000); the category of coal
is thus unknown and unpredictable during the combu stion. Unpredictable change
of coal type may cause the furnace to depart from its normal combustion states,
and in extreme conditions can result in a vicious accident such as burn-out of the
flame in the boiler, ‘‘fire a gun,’’ and hearth exploration, etc.
In recent studies, the X-ray fluorescence (XRF) (Huggins, 2002), pulsed neu-
trons analysis (Belbot et al., 2001), laser ablation inductively coupled mass spec-
trometry (LA-ICP-MS) (Kleiber et al., 2002), near-infrared (NIR) spectroscopy
(Andre
´
s and Bona, 2006) techniques and so on have been used as the coal analyze r
to measure the chemical elements contained in the coal. Because they are very expens-
ive and require complicated installations, they are seldom used. Therefore, online,
in-time, and easy-to-use coal type identification at a power station where a wide range
of coals is needed can help effectively control the normal running of the boiler, and
hence increase the combust ion efficiency and reduce the pollutant emissions.
A flame of coal combustion in the boiler oscillates violently, and the oscillating
features of the flame across its radiation spectra from the infrared (IR) through vis-
ible to the ultraviolet (UV) regions are different at the root area from coal to coal.
The oscillatory characteristics of a flame are often used to determine the failure
detection of the flame (Jones, 1988; Odgaard and Trangbaek, 2006) and the under-
standing of flame structure and stability (Lu et al., 2006) in the furnace. Xu et al.
(2004, 2005) made use of optical sensing technique to identify the fuel type. Flame
oscillating signal s covering the IR, visible, and UV bands can be acqu ired by a
low-cost three-cell flame detector (Xu and Yan, 2006) that is installed in the same
place as a traditional flame-eye and stares at the root area of the flame. The flame
oscillation signals captured by the flame detector can not only be used to monitor
the state of coal combustion, but also to identify the type of coal.
Too many features may make the identification process too complicated and
lead to incorrect prediction. This problem can be avoided by extracting only the rel-
evant features and obtaining new features containing the maximal information from
the original features. The former is called feature selection, while the latter is called
feature dimension reduction (Kwak et al., 2001). As two powerful mathematical
tools, the principal components analysis (PCA) and independent component analysis
(ICA) techniques were introduced for reducing data dimension. PCA uses a set of
basis functions to optimally model the data in the sense of minimum error. The
ICA method can be considered as a generalization of PCA and can find a linear
transform for the observed data by using a set of basis functions where the compo-
nents are not only decorrelated but also as mutually independent as possible
(Widodo and Yang, 2007). In the early literatu re, PCA was used for feature
reduction to minimize the complexity, and fuzzy logic (Xu et al., 2004) and neural
network (Xu et al., 2005) techniques were used to map the featu res of flame oscil-
lation to the type of co al, respectively. Experimental results showed that the neural
network with a preceding step to implement PCA is better than the fuzzy logic
method without PCA in term s of success rate of coal type identification. Cao et al.
(2003) compared PCA, kernel principal component analysis (KPCA), and ICA for
dimensionality reduction in support vector machine by examining the sunspot data,
Santa Fe data set A, and five real futures contracts. The experimental results showed
278 C. TAN ET AL.
Downloaded by [Beihang University] at 16:02 27 February 2012