Professional Documents
Culture Documents
2007 Al 3589 Quantitative Analysis of Formaldehyde
2007 Al 3589 Quantitative Analysis of Formaldehyde
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/244601474
CITATIONS READS
3 1,031
3 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Han Chern Loh on 07 February 2015.
The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
QUANTITATIVE ANALYSIS OF FORMALDEHYDE
USING UV-VIS SPECTROMETER PATTERN
RECOGNITION AND ARTIFICIAL
NEURAL NETWORK
1. Introduction
O H H O
CH3 CH 3
O
C
C
+ 2 C C H3C C C CH3
H H O NH 2
C C
H H3C N CH3
H
Formaldehid Fluoral-P 3,5-diacetyl-1,4-dihydroluthydine
(Equation 1)
ANNs are computational models that share some of the properties of the brain.
They are parallel information-processing structures that attempt to emulate certain
performance characteristics of the biological neural system. These networks consist of
many simples “units” working in parallel with no central control and learning takes
place by modifying the weights between connections. They process information by
their dynamic state response to external inputs [7]. The basic components of an ANNs
are “neurons”, weights and learning rules [8]. Thus, ANNs are also described as data
processing systems that simulate the human brain by building on information through
“learning” [9].
2. Experimental
All reagents used in this study were of analytical-reagent grade. Acetic acid
and acetyl acetone were purchased from J. T. Baker whilst ammonium acetate,
potassium chloride, hydrochloric, potassium hydrogen phthalate, potassium
dihydrogen phosphate, sodium oxide, Borax and sodium bicarbonate were purchased
from Fluka. Formaldehyde was obtained from AnalaR. All chemicals were used
without further purification. Deionised water obtained from Maxima Ultra Pure Water
Elga (Elga Stat Maxima UF model) was used throughout the experiment for solution
preparation. All spectrophotometric measurements were made with Varian Cary 50
UV-Vis Spectrophotometer (Varian Australia Pty. Ltd., Victoria, Australia) using
quartz cell.
Fluoral-P reagent was prepared by dissolving 7.7 g ammonium acetate with deionised
water in 250 mL volumetric flask. This solution was then added with 1.5 mL of 1.0 M
acetic acid and 1.0 mL of 1.0 M acetyl acetone. The mixture was added with
deionised water until the line and was shaken. Buffer solutions were prepared as
shown in Table 1 [37].
2.3 Procedures
The ANN training and data treatment were realised using a Matlab program.
In this study, a three layers feed-forward ANN with a BP training algorithm was
employed for data treatment. The input layer consists of 23 neurons, which
represented the absorption intensities measured at 7 different wavelengths from each
spectrum. The network was trained up to 40,000 cycles and the progress of sum-
squared error (SSE) between the calculated and the measured output was recorded. In
order to check its forecast capability and precision, a new set of input data was
introduced. The best network chosen was based on several tests using the trained
network that incorporated the inspection for data training fitting errors and prediction
test of errors.
The absorption was measured at wavelength of 411 nm for the complex. This
wavelength is very similar with the previous work of 412 nm [5]. The optimum
response of the complex was obtained at pH 7 (Fig. 2). Hence this pH was used
throughout this study. The photostability of this complex is more than 8 hours. The
study of response time was carried out by using three different concentrations of
formaldehyde, i. e. 20, 40 and 60 ppm (Fig. 3). The high reaction rate was obtained
for the first 20 minutes. After 20 minutes, the reaction rate had become lower and
constant.
The effect of the reagent concentration on the complex was studied by using
different initial concentrations of the reagent. Fig. 4 shows the effect of reagent
concentration of Fluoral-P used during complexation. As the concentration of Fluoral-
P is increased, the measured absorbance is also increased. This is due to the more
Fluoral-P-formaldehyde complex has been formed when more reagent is available.
The curvature at higher concentration is expected due to the reagent has been fully
reacts with formaldehyde. The same curvatures have been reported by Loh and co-
worker [27-28,38] as well as Musa and Narayanaswamy works [39]. Reagent
concentration of 1.24 x 10-5 M was chosen for the following study in this study.
In this study, 7 wavelengths points (238, 254, 262, 277, 316, 411 and 460 nm)
from each spectrum were selected to represent the input data for the ANN. These
points were selected due to the general outline of the original spectra were represented
and the variations in the sensor response were significant. The same training method
has been used and reported [27-30] to avoid several problems during network
training. These problems are long training period and large matrices entailed for the
network connection [27-28, 38].
A total of 22 spectra (0.6, 1.2, 2.4, 3.0, 4.2, 4.8, 6.0, 6.6, 7.8, 8.4, 9.6, 10.2,
11.4, 12.0, 13.2, 13.8, 15.0, 15.6, 16.8, 17.4, 18.4 and 19.2 ppm) were used for ANN
training. The optimisation of network was performed by changing the number of
hidden neurons and cycle number (Table 1). The SSE was measured at the end of
each training and being recorded. Table 3 shows the SSE values of network after
undergone different cycle number and hidden neurons. Network with 23 hidden
neurons and 40000 cycle number was chosen for this study due to the network gave
the lowest average calibration error value.
All of these networks were presented with eleven calibration spectra (1.8, 3.6,
5.4, 7.2, 9.0, 10.8, 12.6, 14.4, 16.2, 18.6 and 19.8 ppm) in order to improve the
process in choosing the best architecture of network and to establish their prediction
capability. Table 2 displays the predicted values of all formaldehyde concentrations
with the measured formaldehyde concentrations with the network training. The useful
dynamic range is now extended to a wider range of 0 – 19.8 ppm with average
calibration error and SSE are 0.1771 and 0.5847 respectively.
4. Conclusion
The studies carried out in this work indicate that formaldehyde determination
by using Fluoral-P as reagent is optimum at pH 7 and the 1.24 x 10-5 M reagent
concentration. A good reproducibility (2.12 %) of measurement was obtained. This
study can be developed further as into formaldehyde sensor fabrication. The network,
trained using BP algorithm was highly accurate in predicting the response of this
sensor, with average prediction error of 0.5847 g/L. The ultimate effect of
implementing ANN on formaldehyde detection was broadening the useful response
range from limited linear response (0.60 – 3.60 ppm) to extensive calibration response
(0.00 – 3.60 ppm).
Acknowledgements
SSE 0.5847
Table 3: SSE values of network after undergone different cycle number and hidden
neuron
Wavelength
1.6
1.5
1.4
Abs
1.3
1.2
1.1
1
0 2 4 6 8 10 12 14
pH
Time (min)
0.5
0.4
0.3
Abs
0.2
0.1
0
0 0.1 0.2 0.3
Fluoral-P Concentration (mM)
0.5
0.4
Abs
0.3
0.2
0.1
0
0 2 4 6 8 10 12 14 16 18 20
Formadehyde concentration (ppm)
0.30
0.25
0.20
Abs
0.15
0.10
0.05
0.00
1 2 3 4 5 6 7 8 9 10
Measurement
20ppm 40ppm
1.0
0.8
0.6
Abs
0.4
0.2
0
19.0
230 9.50
360 Formaldehyde
Wavelength 500 4.75 concentration
0 (ppm)